ext4: flush delalloc blocks when space is low
[linux-2.6.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  Goal-directed block allocation by Stephen Tweedie
16  *      (sct@redhat.com), 1993, 1998
17  *  Big-endian to little-endian byte-swapping/bitmaps by
18  *        David S. Miller (davem@caip.rutgers.edu), 1995
19  *  64-bit file support on 64-bit platforms by Jakub Jelinek
20  *      (jj@sunsite.ms.mff.cuni.cz)
21  *
22  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
23  */
24
25 #include <linux/module.h>
26 #include <linux/fs.h>
27 #include <linux/time.h>
28 #include <linux/jbd2.h>
29 #include <linux/highuid.h>
30 #include <linux/pagemap.h>
31 #include <linux/quotaops.h>
32 #include <linux/string.h>
33 #include <linux/buffer_head.h>
34 #include <linux/writeback.h>
35 #include <linux/pagevec.h>
36 #include <linux/mpage.h>
37 #include <linux/namei.h>
38 #include <linux/uio.h>
39 #include <linux/bio.h>
40 #include <linux/workqueue.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "ext4_extents.h"
46
47 #include <trace/events/ext4.h>
48
49 #define MPAGE_DA_EXTENT_TAIL 0x01
50
51 static inline int ext4_begin_ordered_truncate(struct inode *inode,
52                                               loff_t new_size)
53 {
54         return jbd2_journal_begin_ordered_truncate(
55                                         EXT4_SB(inode->i_sb)->s_journal,
56                                         &EXT4_I(inode)->jinode,
57                                         new_size);
58 }
59
60 static void ext4_invalidatepage(struct page *page, unsigned long offset);
61
62 /*
63  * Test whether an inode is a fast symlink.
64  */
65 static int ext4_inode_is_fast_symlink(struct inode *inode)
66 {
67         int ea_blocks = EXT4_I(inode)->i_file_acl ?
68                 (inode->i_sb->s_blocksize >> 9) : 0;
69
70         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
71 }
72
73 /*
74  * Work out how many blocks we need to proceed with the next chunk of a
75  * truncate transaction.
76  */
77 static unsigned long blocks_for_truncate(struct inode *inode)
78 {
79         ext4_lblk_t needed;
80
81         needed = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9);
82
83         /* Give ourselves just enough room to cope with inodes in which
84          * i_blocks is corrupt: we've seen disk corruptions in the past
85          * which resulted in random data in an inode which looked enough
86          * like a regular file for ext4 to try to delete it.  Things
87          * will go a bit crazy if that happens, but at least we should
88          * try not to panic the whole kernel. */
89         if (needed < 2)
90                 needed = 2;
91
92         /* But we need to bound the transaction so we don't overflow the
93          * journal. */
94         if (needed > EXT4_MAX_TRANS_DATA)
95                 needed = EXT4_MAX_TRANS_DATA;
96
97         return EXT4_DATA_TRANS_BLOCKS(inode->i_sb) + needed;
98 }
99
100 /*
101  * Truncate transactions can be complex and absolutely huge.  So we need to
102  * be able to restart the transaction at a conventient checkpoint to make
103  * sure we don't overflow the journal.
104  *
105  * start_transaction gets us a new handle for a truncate transaction,
106  * and extend_transaction tries to extend the existing one a bit.  If
107  * extend fails, we need to propagate the failure up and restart the
108  * transaction in the top-level truncate loop. --sct
109  */
110 static handle_t *start_transaction(struct inode *inode)
111 {
112         handle_t *result;
113
114         result = ext4_journal_start(inode, blocks_for_truncate(inode));
115         if (!IS_ERR(result))
116                 return result;
117
118         ext4_std_error(inode->i_sb, PTR_ERR(result));
119         return result;
120 }
121
122 /*
123  * Try to extend this transaction for the purposes of truncation.
124  *
125  * Returns 0 if we managed to create more room.  If we can't create more
126  * room, and the transaction must be restarted we return 1.
127  */
128 static int try_to_extend_transaction(handle_t *handle, struct inode *inode)
129 {
130         if (!ext4_handle_valid(handle))
131                 return 0;
132         if (ext4_handle_has_enough_credits(handle, EXT4_RESERVE_TRANS_BLOCKS+1))
133                 return 0;
134         if (!ext4_journal_extend(handle, blocks_for_truncate(inode)))
135                 return 0;
136         return 1;
137 }
138
139 /*
140  * Restart the transaction associated with *handle.  This does a commit,
141  * so before we call here everything must be consistently dirtied against
142  * this transaction.
143  */
144 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
145                                  int nblocks)
146 {
147         int ret;
148
149         /*
150          * Drop i_data_sem to avoid deadlock with ext4_get_blocks At this
151          * moment, get_block can be called only for blocks inside i_size since
152          * page cache has been already dropped and writes are blocked by
153          * i_mutex. So we can safely drop the i_data_sem here.
154          */
155         BUG_ON(EXT4_JOURNAL(inode) == NULL);
156         jbd_debug(2, "restarting handle %p\n", handle);
157         up_write(&EXT4_I(inode)->i_data_sem);
158         ret = ext4_journal_restart(handle, blocks_for_truncate(inode));
159         down_write(&EXT4_I(inode)->i_data_sem);
160         ext4_discard_preallocations(inode);
161
162         return ret;
163 }
164
165 /*
166  * Called at the last iput() if i_nlink is zero.
167  */
168 void ext4_delete_inode(struct inode *inode)
169 {
170         handle_t *handle;
171         int err;
172
173         if (ext4_should_order_data(inode))
174                 ext4_begin_ordered_truncate(inode, 0);
175         truncate_inode_pages(&inode->i_data, 0);
176
177         if (is_bad_inode(inode))
178                 goto no_delete;
179
180         handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
181         if (IS_ERR(handle)) {
182                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
183                 /*
184                  * If we're going to skip the normal cleanup, we still need to
185                  * make sure that the in-core orphan linked list is properly
186                  * cleaned up.
187                  */
188                 ext4_orphan_del(NULL, inode);
189                 goto no_delete;
190         }
191
192         if (IS_SYNC(inode))
193                 ext4_handle_sync(handle);
194         inode->i_size = 0;
195         err = ext4_mark_inode_dirty(handle, inode);
196         if (err) {
197                 ext4_warning(inode->i_sb, __func__,
198                              "couldn't mark inode dirty (err %d)", err);
199                 goto stop_handle;
200         }
201         if (inode->i_blocks)
202                 ext4_truncate(inode);
203
204         /*
205          * ext4_ext_truncate() doesn't reserve any slop when it
206          * restarts journal transactions; therefore there may not be
207          * enough credits left in the handle to remove the inode from
208          * the orphan list and set the dtime field.
209          */
210         if (!ext4_handle_has_enough_credits(handle, 3)) {
211                 err = ext4_journal_extend(handle, 3);
212                 if (err > 0)
213                         err = ext4_journal_restart(handle, 3);
214                 if (err != 0) {
215                         ext4_warning(inode->i_sb, __func__,
216                                      "couldn't extend journal (err %d)", err);
217                 stop_handle:
218                         ext4_journal_stop(handle);
219                         goto no_delete;
220                 }
221         }
222
223         /*
224          * Kill off the orphan record which ext4_truncate created.
225          * AKPM: I think this can be inside the above `if'.
226          * Note that ext4_orphan_del() has to be able to cope with the
227          * deletion of a non-existent orphan - this is because we don't
228          * know if ext4_truncate() actually created an orphan record.
229          * (Well, we could do this if we need to, but heck - it works)
230          */
231         ext4_orphan_del(handle, inode);
232         EXT4_I(inode)->i_dtime  = get_seconds();
233
234         /*
235          * One subtle ordering requirement: if anything has gone wrong
236          * (transaction abort, IO errors, whatever), then we can still
237          * do these next steps (the fs will already have been marked as
238          * having errors), but we can't free the inode if the mark_dirty
239          * fails.
240          */
241         if (ext4_mark_inode_dirty(handle, inode))
242                 /* If that failed, just do the required in-core inode clear. */
243                 clear_inode(inode);
244         else
245                 ext4_free_inode(handle, inode);
246         ext4_journal_stop(handle);
247         return;
248 no_delete:
249         clear_inode(inode);     /* We must guarantee clearing of inode... */
250 }
251
252 typedef struct {
253         __le32  *p;
254         __le32  key;
255         struct buffer_head *bh;
256 } Indirect;
257
258 static inline void add_chain(Indirect *p, struct buffer_head *bh, __le32 *v)
259 {
260         p->key = *(p->p = v);
261         p->bh = bh;
262 }
263
264 /**
265  *      ext4_block_to_path - parse the block number into array of offsets
266  *      @inode: inode in question (we are only interested in its superblock)
267  *      @i_block: block number to be parsed
268  *      @offsets: array to store the offsets in
269  *      @boundary: set this non-zero if the referred-to block is likely to be
270  *             followed (on disk) by an indirect block.
271  *
272  *      To store the locations of file's data ext4 uses a data structure common
273  *      for UNIX filesystems - tree of pointers anchored in the inode, with
274  *      data blocks at leaves and indirect blocks in intermediate nodes.
275  *      This function translates the block number into path in that tree -
276  *      return value is the path length and @offsets[n] is the offset of
277  *      pointer to (n+1)th node in the nth one. If @block is out of range
278  *      (negative or too large) warning is printed and zero returned.
279  *
280  *      Note: function doesn't find node addresses, so no IO is needed. All
281  *      we need to know is the capacity of indirect blocks (taken from the
282  *      inode->i_sb).
283  */
284
285 /*
286  * Portability note: the last comparison (check that we fit into triple
287  * indirect block) is spelled differently, because otherwise on an
288  * architecture with 32-bit longs and 8Kb pages we might get into trouble
289  * if our filesystem had 8Kb blocks. We might use long long, but that would
290  * kill us on x86. Oh, well, at least the sign propagation does not matter -
291  * i_block would have to be negative in the very beginning, so we would not
292  * get there at all.
293  */
294
295 static int ext4_block_to_path(struct inode *inode,
296                               ext4_lblk_t i_block,
297                               ext4_lblk_t offsets[4], int *boundary)
298 {
299         int ptrs = EXT4_ADDR_PER_BLOCK(inode->i_sb);
300         int ptrs_bits = EXT4_ADDR_PER_BLOCK_BITS(inode->i_sb);
301         const long direct_blocks = EXT4_NDIR_BLOCKS,
302                 indirect_blocks = ptrs,
303                 double_blocks = (1 << (ptrs_bits * 2));
304         int n = 0;
305         int final = 0;
306
307         if (i_block < direct_blocks) {
308                 offsets[n++] = i_block;
309                 final = direct_blocks;
310         } else if ((i_block -= direct_blocks) < indirect_blocks) {
311                 offsets[n++] = EXT4_IND_BLOCK;
312                 offsets[n++] = i_block;
313                 final = ptrs;
314         } else if ((i_block -= indirect_blocks) < double_blocks) {
315                 offsets[n++] = EXT4_DIND_BLOCK;
316                 offsets[n++] = i_block >> ptrs_bits;
317                 offsets[n++] = i_block & (ptrs - 1);
318                 final = ptrs;
319         } else if (((i_block -= double_blocks) >> (ptrs_bits * 2)) < ptrs) {
320                 offsets[n++] = EXT4_TIND_BLOCK;
321                 offsets[n++] = i_block >> (ptrs_bits * 2);
322                 offsets[n++] = (i_block >> ptrs_bits) & (ptrs - 1);
323                 offsets[n++] = i_block & (ptrs - 1);
324                 final = ptrs;
325         } else {
326                 ext4_warning(inode->i_sb, "ext4_block_to_path",
327                              "block %lu > max in inode %lu",
328                              i_block + direct_blocks +
329                              indirect_blocks + double_blocks, inode->i_ino);
330         }
331         if (boundary)
332                 *boundary = final - 1 - (i_block & (ptrs - 1));
333         return n;
334 }
335
336 static int __ext4_check_blockref(const char *function, struct inode *inode,
337                                  __le32 *p, unsigned int max)
338 {
339         __le32 *bref = p;
340         unsigned int blk;
341
342         while (bref < p+max) {
343                 blk = le32_to_cpu(*bref++);
344                 if (blk &&
345                     unlikely(!ext4_data_block_valid(EXT4_SB(inode->i_sb),
346                                                     blk, 1))) {
347                         ext4_error(inode->i_sb, function,
348                                    "invalid block reference %u "
349                                    "in inode #%lu", blk, inode->i_ino);
350                         return -EIO;
351                 }
352         }
353         return 0;
354 }
355
356
357 #define ext4_check_indirect_blockref(inode, bh)                         \
358         __ext4_check_blockref(__func__, inode, (__le32 *)(bh)->b_data,  \
359                               EXT4_ADDR_PER_BLOCK((inode)->i_sb))
360
361 #define ext4_check_inode_blockref(inode)                                \
362         __ext4_check_blockref(__func__, inode, EXT4_I(inode)->i_data,   \
363                               EXT4_NDIR_BLOCKS)
364
365 /**
366  *      ext4_get_branch - read the chain of indirect blocks leading to data
367  *      @inode: inode in question
368  *      @depth: depth of the chain (1 - direct pointer, etc.)
369  *      @offsets: offsets of pointers in inode/indirect blocks
370  *      @chain: place to store the result
371  *      @err: here we store the error value
372  *
373  *      Function fills the array of triples <key, p, bh> and returns %NULL
374  *      if everything went OK or the pointer to the last filled triple
375  *      (incomplete one) otherwise. Upon the return chain[i].key contains
376  *      the number of (i+1)-th block in the chain (as it is stored in memory,
377  *      i.e. little-endian 32-bit), chain[i].p contains the address of that
378  *      number (it points into struct inode for i==0 and into the bh->b_data
379  *      for i>0) and chain[i].bh points to the buffer_head of i-th indirect
380  *      block for i>0 and NULL for i==0. In other words, it holds the block
381  *      numbers of the chain, addresses they were taken from (and where we can
382  *      verify that chain did not change) and buffer_heads hosting these
383  *      numbers.
384  *
385  *      Function stops when it stumbles upon zero pointer (absent block)
386  *              (pointer to last triple returned, *@err == 0)
387  *      or when it gets an IO error reading an indirect block
388  *              (ditto, *@err == -EIO)
389  *      or when it reads all @depth-1 indirect blocks successfully and finds
390  *      the whole chain, all way to the data (returns %NULL, *err == 0).
391  *
392  *      Need to be called with
393  *      down_read(&EXT4_I(inode)->i_data_sem)
394  */
395 static Indirect *ext4_get_branch(struct inode *inode, int depth,
396                                  ext4_lblk_t  *offsets,
397                                  Indirect chain[4], int *err)
398 {
399         struct super_block *sb = inode->i_sb;
400         Indirect *p = chain;
401         struct buffer_head *bh;
402
403         *err = 0;
404         /* i_data is not going away, no lock needed */
405         add_chain(chain, NULL, EXT4_I(inode)->i_data + *offsets);
406         if (!p->key)
407                 goto no_block;
408         while (--depth) {
409                 bh = sb_getblk(sb, le32_to_cpu(p->key));
410                 if (unlikely(!bh))
411                         goto failure;
412
413                 if (!bh_uptodate_or_lock(bh)) {
414                         if (bh_submit_read(bh) < 0) {
415                                 put_bh(bh);
416                                 goto failure;
417                         }
418                         /* validate block references */
419                         if (ext4_check_indirect_blockref(inode, bh)) {
420                                 put_bh(bh);
421                                 goto failure;
422                         }
423                 }
424
425                 add_chain(++p, bh, (__le32 *)bh->b_data + *++offsets);
426                 /* Reader: end */
427                 if (!p->key)
428                         goto no_block;
429         }
430         return NULL;
431
432 failure:
433         *err = -EIO;
434 no_block:
435         return p;
436 }
437
438 /**
439  *      ext4_find_near - find a place for allocation with sufficient locality
440  *      @inode: owner
441  *      @ind: descriptor of indirect block.
442  *
443  *      This function returns the preferred place for block allocation.
444  *      It is used when heuristic for sequential allocation fails.
445  *      Rules are:
446  *        + if there is a block to the left of our position - allocate near it.
447  *        + if pointer will live in indirect block - allocate near that block.
448  *        + if pointer will live in inode - allocate in the same
449  *          cylinder group.
450  *
451  * In the latter case we colour the starting block by the callers PID to
452  * prevent it from clashing with concurrent allocations for a different inode
453  * in the same block group.   The PID is used here so that functionally related
454  * files will be close-by on-disk.
455  *
456  *      Caller must make sure that @ind is valid and will stay that way.
457  */
458 static ext4_fsblk_t ext4_find_near(struct inode *inode, Indirect *ind)
459 {
460         struct ext4_inode_info *ei = EXT4_I(inode);
461         __le32 *start = ind->bh ? (__le32 *) ind->bh->b_data : ei->i_data;
462         __le32 *p;
463         ext4_fsblk_t bg_start;
464         ext4_fsblk_t last_block;
465         ext4_grpblk_t colour;
466         ext4_group_t block_group;
467         int flex_size = ext4_flex_bg_size(EXT4_SB(inode->i_sb));
468
469         /* Try to find previous block */
470         for (p = ind->p - 1; p >= start; p--) {
471                 if (*p)
472                         return le32_to_cpu(*p);
473         }
474
475         /* No such thing, so let's try location of indirect block */
476         if (ind->bh)
477                 return ind->bh->b_blocknr;
478
479         /*
480          * It is going to be referred to from the inode itself? OK, just put it
481          * into the same cylinder group then.
482          */
483         block_group = ei->i_block_group;
484         if (flex_size >= EXT4_FLEX_SIZE_DIR_ALLOC_SCHEME) {
485                 block_group &= ~(flex_size-1);
486                 if (S_ISREG(inode->i_mode))
487                         block_group++;
488         }
489         bg_start = ext4_group_first_block_no(inode->i_sb, block_group);
490         last_block = ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es) - 1;
491
492         /*
493          * If we are doing delayed allocation, we don't need take
494          * colour into account.
495          */
496         if (test_opt(inode->i_sb, DELALLOC))
497                 return bg_start;
498
499         if (bg_start + EXT4_BLOCKS_PER_GROUP(inode->i_sb) <= last_block)
500                 colour = (current->pid % 16) *
501                         (EXT4_BLOCKS_PER_GROUP(inode->i_sb) / 16);
502         else
503                 colour = (current->pid % 16) * ((last_block - bg_start) / 16);
504         return bg_start + colour;
505 }
506
507 /**
508  *      ext4_find_goal - find a preferred place for allocation.
509  *      @inode: owner
510  *      @block:  block we want
511  *      @partial: pointer to the last triple within a chain
512  *
513  *      Normally this function find the preferred place for block allocation,
514  *      returns it.
515  *      Because this is only used for non-extent files, we limit the block nr
516  *      to 32 bits.
517  */
518 static ext4_fsblk_t ext4_find_goal(struct inode *inode, ext4_lblk_t block,
519                                    Indirect *partial)
520 {
521         ext4_fsblk_t goal;
522
523         /*
524          * XXX need to get goal block from mballoc's data structures
525          */
526
527         goal = ext4_find_near(inode, partial);
528         goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
529         return goal;
530 }
531
532 /**
533  *      ext4_blks_to_allocate: Look up the block map and count the number
534  *      of direct blocks need to be allocated for the given branch.
535  *
536  *      @branch: chain of indirect blocks
537  *      @k: number of blocks need for indirect blocks
538  *      @blks: number of data blocks to be mapped.
539  *      @blocks_to_boundary:  the offset in the indirect block
540  *
541  *      return the total number of blocks to be allocate, including the
542  *      direct and indirect blocks.
543  */
544 static int ext4_blks_to_allocate(Indirect *branch, int k, unsigned int blks,
545                                  int blocks_to_boundary)
546 {
547         unsigned int count = 0;
548
549         /*
550          * Simple case, [t,d]Indirect block(s) has not allocated yet
551          * then it's clear blocks on that path have not allocated
552          */
553         if (k > 0) {
554                 /* right now we don't handle cross boundary allocation */
555                 if (blks < blocks_to_boundary + 1)
556                         count += blks;
557                 else
558                         count += blocks_to_boundary + 1;
559                 return count;
560         }
561
562         count++;
563         while (count < blks && count <= blocks_to_boundary &&
564                 le32_to_cpu(*(branch[0].p + count)) == 0) {
565                 count++;
566         }
567         return count;
568 }
569
570 /**
571  *      ext4_alloc_blocks: multiple allocate blocks needed for a branch
572  *      @indirect_blks: the number of blocks need to allocate for indirect
573  *                      blocks
574  *
575  *      @new_blocks: on return it will store the new block numbers for
576  *      the indirect blocks(if needed) and the first direct block,
577  *      @blks:  on return it will store the total number of allocated
578  *              direct blocks
579  */
580 static int ext4_alloc_blocks(handle_t *handle, struct inode *inode,
581                              ext4_lblk_t iblock, ext4_fsblk_t goal,
582                              int indirect_blks, int blks,
583                              ext4_fsblk_t new_blocks[4], int *err)
584 {
585         struct ext4_allocation_request ar;
586         int target, i;
587         unsigned long count = 0, blk_allocated = 0;
588         int index = 0;
589         ext4_fsblk_t current_block = 0;
590         int ret = 0;
591
592         /*
593          * Here we try to allocate the requested multiple blocks at once,
594          * on a best-effort basis.
595          * To build a branch, we should allocate blocks for
596          * the indirect blocks(if not allocated yet), and at least
597          * the first direct block of this branch.  That's the
598          * minimum number of blocks need to allocate(required)
599          */
600         /* first we try to allocate the indirect blocks */
601         target = indirect_blks;
602         while (target > 0) {
603                 count = target;
604                 /* allocating blocks for indirect blocks and direct blocks */
605                 current_block = ext4_new_meta_blocks(handle, inode,
606                                                         goal, &count, err);
607                 if (*err)
608                         goto failed_out;
609
610                 BUG_ON(current_block + count > EXT4_MAX_BLOCK_FILE_PHYS);
611
612                 target -= count;
613                 /* allocate blocks for indirect blocks */
614                 while (index < indirect_blks && count) {
615                         new_blocks[index++] = current_block++;
616                         count--;
617                 }
618                 if (count > 0) {
619                         /*
620                          * save the new block number
621                          * for the first direct block
622                          */
623                         new_blocks[index] = current_block;
624                         printk(KERN_INFO "%s returned more blocks than "
625                                                 "requested\n", __func__);
626                         WARN_ON(1);
627                         break;
628                 }
629         }
630
631         target = blks - count ;
632         blk_allocated = count;
633         if (!target)
634                 goto allocated;
635         /* Now allocate data blocks */
636         memset(&ar, 0, sizeof(ar));
637         ar.inode = inode;
638         ar.goal = goal;
639         ar.len = target;
640         ar.logical = iblock;
641         if (S_ISREG(inode->i_mode))
642                 /* enable in-core preallocation only for regular files */
643                 ar.flags = EXT4_MB_HINT_DATA;
644
645         current_block = ext4_mb_new_blocks(handle, &ar, err);
646         BUG_ON(current_block + ar.len > EXT4_MAX_BLOCK_FILE_PHYS);
647
648         if (*err && (target == blks)) {
649                 /*
650                  * if the allocation failed and we didn't allocate
651                  * any blocks before
652                  */
653                 goto failed_out;
654         }
655         if (!*err) {
656                 if (target == blks) {
657                         /*
658                          * save the new block number
659                          * for the first direct block
660                          */
661                         new_blocks[index] = current_block;
662                 }
663                 blk_allocated += ar.len;
664         }
665 allocated:
666         /* total number of blocks allocated for direct blocks */
667         ret = blk_allocated;
668         *err = 0;
669         return ret;
670 failed_out:
671         for (i = 0; i < index; i++)
672                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
673         return ret;
674 }
675
676 /**
677  *      ext4_alloc_branch - allocate and set up a chain of blocks.
678  *      @inode: owner
679  *      @indirect_blks: number of allocated indirect blocks
680  *      @blks: number of allocated direct blocks
681  *      @offsets: offsets (in the blocks) to store the pointers to next.
682  *      @branch: place to store the chain in.
683  *
684  *      This function allocates blocks, zeroes out all but the last one,
685  *      links them into chain and (if we are synchronous) writes them to disk.
686  *      In other words, it prepares a branch that can be spliced onto the
687  *      inode. It stores the information about that chain in the branch[], in
688  *      the same format as ext4_get_branch() would do. We are calling it after
689  *      we had read the existing part of chain and partial points to the last
690  *      triple of that (one with zero ->key). Upon the exit we have the same
691  *      picture as after the successful ext4_get_block(), except that in one
692  *      place chain is disconnected - *branch->p is still zero (we did not
693  *      set the last link), but branch->key contains the number that should
694  *      be placed into *branch->p to fill that gap.
695  *
696  *      If allocation fails we free all blocks we've allocated (and forget
697  *      their buffer_heads) and return the error value the from failed
698  *      ext4_alloc_block() (normally -ENOSPC). Otherwise we set the chain
699  *      as described above and return 0.
700  */
701 static int ext4_alloc_branch(handle_t *handle, struct inode *inode,
702                              ext4_lblk_t iblock, int indirect_blks,
703                              int *blks, ext4_fsblk_t goal,
704                              ext4_lblk_t *offsets, Indirect *branch)
705 {
706         int blocksize = inode->i_sb->s_blocksize;
707         int i, n = 0;
708         int err = 0;
709         struct buffer_head *bh;
710         int num;
711         ext4_fsblk_t new_blocks[4];
712         ext4_fsblk_t current_block;
713
714         num = ext4_alloc_blocks(handle, inode, iblock, goal, indirect_blks,
715                                 *blks, new_blocks, &err);
716         if (err)
717                 return err;
718
719         branch[0].key = cpu_to_le32(new_blocks[0]);
720         /*
721          * metadata blocks and data blocks are allocated.
722          */
723         for (n = 1; n <= indirect_blks;  n++) {
724                 /*
725                  * Get buffer_head for parent block, zero it out
726                  * and set the pointer to new one, then send
727                  * parent to disk.
728                  */
729                 bh = sb_getblk(inode->i_sb, new_blocks[n-1]);
730                 branch[n].bh = bh;
731                 lock_buffer(bh);
732                 BUFFER_TRACE(bh, "call get_create_access");
733                 err = ext4_journal_get_create_access(handle, bh);
734                 if (err) {
735                         /* Don't brelse(bh) here; it's done in
736                          * ext4_journal_forget() below */
737                         unlock_buffer(bh);
738                         goto failed;
739                 }
740
741                 memset(bh->b_data, 0, blocksize);
742                 branch[n].p = (__le32 *) bh->b_data + offsets[n];
743                 branch[n].key = cpu_to_le32(new_blocks[n]);
744                 *branch[n].p = branch[n].key;
745                 if (n == indirect_blks) {
746                         current_block = new_blocks[n];
747                         /*
748                          * End of chain, update the last new metablock of
749                          * the chain to point to the new allocated
750                          * data blocks numbers
751                          */
752                         for (i = 1; i < num; i++)
753                                 *(branch[n].p + i) = cpu_to_le32(++current_block);
754                 }
755                 BUFFER_TRACE(bh, "marking uptodate");
756                 set_buffer_uptodate(bh);
757                 unlock_buffer(bh);
758
759                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
760                 err = ext4_handle_dirty_metadata(handle, inode, bh);
761                 if (err)
762                         goto failed;
763         }
764         *blks = num;
765         return err;
766 failed:
767         /* Allocation failed, free what we already allocated */
768         ext4_free_blocks(handle, inode, 0, new_blocks[0], 1, 0);
769         for (i = 1; i <= n ; i++) {
770                 /* 
771                  * branch[i].bh is newly allocated, so there is no
772                  * need to revoke the block, which is why we don't
773                  * need to set EXT4_FREE_BLOCKS_METADATA.
774                  */
775                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1,
776                                  EXT4_FREE_BLOCKS_FORGET);
777         }
778         for (i = n+1; i < indirect_blks; i++)
779                 ext4_free_blocks(handle, inode, 0, new_blocks[i], 1, 0);
780
781         ext4_free_blocks(handle, inode, 0, new_blocks[i], num, 0);
782
783         return err;
784 }
785
786 /**
787  * ext4_splice_branch - splice the allocated branch onto inode.
788  * @inode: owner
789  * @block: (logical) number of block we are adding
790  * @chain: chain of indirect blocks (with a missing link - see
791  *      ext4_alloc_branch)
792  * @where: location of missing link
793  * @num:   number of indirect blocks we are adding
794  * @blks:  number of direct blocks we are adding
795  *
796  * This function fills the missing link and does all housekeeping needed in
797  * inode (->i_blocks, etc.). In case of success we end up with the full
798  * chain to new block and return 0.
799  */
800 static int ext4_splice_branch(handle_t *handle, struct inode *inode,
801                               ext4_lblk_t block, Indirect *where, int num,
802                               int blks)
803 {
804         int i;
805         int err = 0;
806         ext4_fsblk_t current_block;
807
808         /*
809          * If we're splicing into a [td]indirect block (as opposed to the
810          * inode) then we need to get write access to the [td]indirect block
811          * before the splice.
812          */
813         if (where->bh) {
814                 BUFFER_TRACE(where->bh, "get_write_access");
815                 err = ext4_journal_get_write_access(handle, where->bh);
816                 if (err)
817                         goto err_out;
818         }
819         /* That's it */
820
821         *where->p = where->key;
822
823         /*
824          * Update the host buffer_head or inode to point to more just allocated
825          * direct blocks blocks
826          */
827         if (num == 0 && blks > 1) {
828                 current_block = le32_to_cpu(where->key) + 1;
829                 for (i = 1; i < blks; i++)
830                         *(where->p + i) = cpu_to_le32(current_block++);
831         }
832
833         /* We are done with atomic stuff, now do the rest of housekeeping */
834         /* had we spliced it onto indirect block? */
835         if (where->bh) {
836                 /*
837                  * If we spliced it onto an indirect block, we haven't
838                  * altered the inode.  Note however that if it is being spliced
839                  * onto an indirect block at the very end of the file (the
840                  * file is growing) then we *will* alter the inode to reflect
841                  * the new i_size.  But that is not done here - it is done in
842                  * generic_commit_write->__mark_inode_dirty->ext4_dirty_inode.
843                  */
844                 jbd_debug(5, "splicing indirect only\n");
845                 BUFFER_TRACE(where->bh, "call ext4_handle_dirty_metadata");
846                 err = ext4_handle_dirty_metadata(handle, inode, where->bh);
847                 if (err)
848                         goto err_out;
849         } else {
850                 /*
851                  * OK, we spliced it into the inode itself on a direct block.
852                  */
853                 ext4_mark_inode_dirty(handle, inode);
854                 jbd_debug(5, "splicing direct\n");
855         }
856         return err;
857
858 err_out:
859         for (i = 1; i <= num; i++) {
860                 /* 
861                  * branch[i].bh is newly allocated, so there is no
862                  * need to revoke the block, which is why we don't
863                  * need to set EXT4_FREE_BLOCKS_METADATA.
864                  */
865                 ext4_free_blocks(handle, inode, where[i].bh, 0, 1,
866                                  EXT4_FREE_BLOCKS_FORGET);
867         }
868         ext4_free_blocks(handle, inode, 0, le32_to_cpu(where[num].key),
869                          blks, 0);
870
871         return err;
872 }
873
874 /*
875  * The ext4_ind_get_blocks() function handles non-extents inodes
876  * (i.e., using the traditional indirect/double-indirect i_blocks
877  * scheme) for ext4_get_blocks().
878  *
879  * Allocation strategy is simple: if we have to allocate something, we will
880  * have to go the whole way to leaf. So let's do it before attaching anything
881  * to tree, set linkage between the newborn blocks, write them if sync is
882  * required, recheck the path, free and repeat if check fails, otherwise
883  * set the last missing link (that will protect us from any truncate-generated
884  * removals - all blocks on the path are immune now) and possibly force the
885  * write on the parent block.
886  * That has a nice additional property: no special recovery from the failed
887  * allocations is needed - we simply release blocks and do not touch anything
888  * reachable from inode.
889  *
890  * `handle' can be NULL if create == 0.
891  *
892  * return > 0, # of blocks mapped or allocated.
893  * return = 0, if plain lookup failed.
894  * return < 0, error case.
895  *
896  * The ext4_ind_get_blocks() function should be called with
897  * down_write(&EXT4_I(inode)->i_data_sem) if allocating filesystem
898  * blocks (i.e., flags has EXT4_GET_BLOCKS_CREATE set) or
899  * down_read(&EXT4_I(inode)->i_data_sem) if not allocating file system
900  * blocks.
901  */
902 static int ext4_ind_get_blocks(handle_t *handle, struct inode *inode,
903                                ext4_lblk_t iblock, unsigned int maxblocks,
904                                struct buffer_head *bh_result,
905                                int flags)
906 {
907         int err = -EIO;
908         ext4_lblk_t offsets[4];
909         Indirect chain[4];
910         Indirect *partial;
911         ext4_fsblk_t goal;
912         int indirect_blks;
913         int blocks_to_boundary = 0;
914         int depth;
915         int count = 0;
916         ext4_fsblk_t first_block = 0;
917
918         J_ASSERT(!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL));
919         J_ASSERT(handle != NULL || (flags & EXT4_GET_BLOCKS_CREATE) == 0);
920         depth = ext4_block_to_path(inode, iblock, offsets,
921                                    &blocks_to_boundary);
922
923         if (depth == 0)
924                 goto out;
925
926         partial = ext4_get_branch(inode, depth, offsets, chain, &err);
927
928         /* Simplest case - block found, no allocation needed */
929         if (!partial) {
930                 first_block = le32_to_cpu(chain[depth - 1].key);
931                 clear_buffer_new(bh_result);
932                 count++;
933                 /*map more blocks*/
934                 while (count < maxblocks && count <= blocks_to_boundary) {
935                         ext4_fsblk_t blk;
936
937                         blk = le32_to_cpu(*(chain[depth-1].p + count));
938
939                         if (blk == first_block + count)
940                                 count++;
941                         else
942                                 break;
943                 }
944                 goto got_it;
945         }
946
947         /* Next simple case - plain lookup or failed read of indirect block */
948         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0 || err == -EIO)
949                 goto cleanup;
950
951         /*
952          * Okay, we need to do block allocation.
953         */
954         goal = ext4_find_goal(inode, iblock, partial);
955
956         /* the number of blocks need to allocate for [d,t]indirect blocks */
957         indirect_blks = (chain + depth) - partial - 1;
958
959         /*
960          * Next look up the indirect map to count the totoal number of
961          * direct blocks to allocate for this branch.
962          */
963         count = ext4_blks_to_allocate(partial, indirect_blks,
964                                         maxblocks, blocks_to_boundary);
965         /*
966          * Block out ext4_truncate while we alter the tree
967          */
968         err = ext4_alloc_branch(handle, inode, iblock, indirect_blks,
969                                 &count, goal,
970                                 offsets + (partial - chain), partial);
971
972         /*
973          * The ext4_splice_branch call will free and forget any buffers
974          * on the new chain if there is a failure, but that risks using
975          * up transaction credits, especially for bitmaps where the
976          * credits cannot be returned.  Can we handle this somehow?  We
977          * may need to return -EAGAIN upwards in the worst case.  --sct
978          */
979         if (!err)
980                 err = ext4_splice_branch(handle, inode, iblock,
981                                          partial, indirect_blks, count);
982         if (err)
983                 goto cleanup;
984
985         set_buffer_new(bh_result);
986
987         ext4_update_inode_fsync_trans(handle, inode, 1);
988 got_it:
989         map_bh(bh_result, inode->i_sb, le32_to_cpu(chain[depth-1].key));
990         if (count > blocks_to_boundary)
991                 set_buffer_boundary(bh_result);
992         err = count;
993         /* Clean up and exit */
994         partial = chain + depth - 1;    /* the whole chain */
995 cleanup:
996         while (partial > chain) {
997                 BUFFER_TRACE(partial->bh, "call brelse");
998                 brelse(partial->bh);
999                 partial--;
1000         }
1001         BUFFER_TRACE(bh_result, "returned");
1002 out:
1003         return err;
1004 }
1005
1006 #ifdef CONFIG_QUOTA
1007 qsize_t *ext4_get_reserved_space(struct inode *inode)
1008 {
1009         return &EXT4_I(inode)->i_reserved_quota;
1010 }
1011 #endif
1012 /*
1013  * Calculate the number of metadata blocks need to reserve
1014  * to allocate @blocks for non extent file based file
1015  */
1016 static int ext4_indirect_calc_metadata_amount(struct inode *inode, int blocks)
1017 {
1018         int icap = EXT4_ADDR_PER_BLOCK(inode->i_sb);
1019         int ind_blks, dind_blks, tind_blks;
1020
1021         /* number of new indirect blocks needed */
1022         ind_blks = (blocks + icap - 1) / icap;
1023
1024         dind_blks = (ind_blks + icap - 1) / icap;
1025
1026         tind_blks = 1;
1027
1028         return ind_blks + dind_blks + tind_blks;
1029 }
1030
1031 /*
1032  * Calculate the number of metadata blocks need to reserve
1033  * to allocate given number of blocks
1034  */
1035 static int ext4_calc_metadata_amount(struct inode *inode, int blocks)
1036 {
1037         if (!blocks)
1038                 return 0;
1039
1040         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
1041                 return ext4_ext_calc_metadata_amount(inode, blocks);
1042
1043         return ext4_indirect_calc_metadata_amount(inode, blocks);
1044 }
1045
1046 static void ext4_da_update_reserve_space(struct inode *inode, int used)
1047 {
1048         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1049         int total, mdb, mdb_free, mdb_claim = 0;
1050
1051         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1052         /* recalculate the number of metablocks still need to be reserved */
1053         total = EXT4_I(inode)->i_reserved_data_blocks - used;
1054         mdb = ext4_calc_metadata_amount(inode, total);
1055
1056         /* figure out how many metablocks to release */
1057         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1058         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1059
1060         if (mdb_free) {
1061                 /* Account for allocated meta_blocks */
1062                 mdb_claim = EXT4_I(inode)->i_allocated_meta_blocks;
1063                 BUG_ON(mdb_free < mdb_claim);
1064                 mdb_free -= mdb_claim;
1065
1066                 /* update fs dirty blocks counter */
1067                 percpu_counter_sub(&sbi->s_dirtyblocks_counter, mdb_free);
1068                 EXT4_I(inode)->i_allocated_meta_blocks = 0;
1069                 EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1070         }
1071
1072         /* update per-inode reservations */
1073         BUG_ON(used  > EXT4_I(inode)->i_reserved_data_blocks);
1074         EXT4_I(inode)->i_reserved_data_blocks -= used;
1075         percpu_counter_sub(&sbi->s_dirtyblocks_counter, used + mdb_claim);
1076         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1077
1078         vfs_dq_claim_block(inode, used + mdb_claim);
1079
1080         /*
1081          * free those over-booking quota for metadata blocks
1082          */
1083         if (mdb_free)
1084                 vfs_dq_release_reservation_block(inode, mdb_free);
1085
1086         /*
1087          * If we have done all the pending block allocations and if
1088          * there aren't any writers on the inode, we can discard the
1089          * inode's preallocations.
1090          */
1091         if (!total && (atomic_read(&inode->i_writecount) == 0))
1092                 ext4_discard_preallocations(inode);
1093 }
1094
1095 static int check_block_validity(struct inode *inode, const char *msg,
1096                                 sector_t logical, sector_t phys, int len)
1097 {
1098         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), phys, len)) {
1099                 ext4_error(inode->i_sb, msg,
1100                            "inode #%lu logical block %llu mapped to %llu "
1101                            "(size %d)", inode->i_ino,
1102                            (unsigned long long) logical,
1103                            (unsigned long long) phys, len);
1104                 return -EIO;
1105         }
1106         return 0;
1107 }
1108
1109 /*
1110  * Return the number of contiguous dirty pages in a given inode
1111  * starting at page frame idx.
1112  */
1113 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
1114                                     unsigned int max_pages)
1115 {
1116         struct address_space *mapping = inode->i_mapping;
1117         pgoff_t index;
1118         struct pagevec pvec;
1119         pgoff_t num = 0;
1120         int i, nr_pages, done = 0;
1121
1122         if (max_pages == 0)
1123                 return 0;
1124         pagevec_init(&pvec, 0);
1125         while (!done) {
1126                 index = idx;
1127                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
1128                                               PAGECACHE_TAG_DIRTY,
1129                                               (pgoff_t)PAGEVEC_SIZE);
1130                 if (nr_pages == 0)
1131                         break;
1132                 for (i = 0; i < nr_pages; i++) {
1133                         struct page *page = pvec.pages[i];
1134                         struct buffer_head *bh, *head;
1135
1136                         lock_page(page);
1137                         if (unlikely(page->mapping != mapping) ||
1138                             !PageDirty(page) ||
1139                             PageWriteback(page) ||
1140                             page->index != idx) {
1141                                 done = 1;
1142                                 unlock_page(page);
1143                                 break;
1144                         }
1145                         if (page_has_buffers(page)) {
1146                                 bh = head = page_buffers(page);
1147                                 do {
1148                                         if (!buffer_delay(bh) &&
1149                                             !buffer_unwritten(bh))
1150                                                 done = 1;
1151                                         bh = bh->b_this_page;
1152                                 } while (!done && (bh != head));
1153                         }
1154                         unlock_page(page);
1155                         if (done)
1156                                 break;
1157                         idx++;
1158                         num++;
1159                         if (num >= max_pages)
1160                                 break;
1161                 }
1162                 pagevec_release(&pvec);
1163         }
1164         return num;
1165 }
1166
1167 /*
1168  * The ext4_get_blocks() function tries to look up the requested blocks,
1169  * and returns if the blocks are already mapped.
1170  *
1171  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
1172  * and store the allocated blocks in the result buffer head and mark it
1173  * mapped.
1174  *
1175  * If file type is extents based, it will call ext4_ext_get_blocks(),
1176  * Otherwise, call with ext4_ind_get_blocks() to handle indirect mapping
1177  * based files
1178  *
1179  * On success, it returns the number of blocks being mapped or allocate.
1180  * if create==0 and the blocks are pre-allocated and uninitialized block,
1181  * the result buffer head is unmapped. If the create ==1, it will make sure
1182  * the buffer head is mapped.
1183  *
1184  * It returns 0 if plain look up failed (blocks have not been allocated), in
1185  * that casem, buffer head is unmapped
1186  *
1187  * It returns the error in case of allocation failure.
1188  */
1189 int ext4_get_blocks(handle_t *handle, struct inode *inode, sector_t block,
1190                     unsigned int max_blocks, struct buffer_head *bh,
1191                     int flags)
1192 {
1193         int retval;
1194
1195         clear_buffer_mapped(bh);
1196         clear_buffer_unwritten(bh);
1197
1198         ext_debug("ext4_get_blocks(): inode %lu, flag %d, max_blocks %u,"
1199                   "logical block %lu\n", inode->i_ino, flags, max_blocks,
1200                   (unsigned long)block);
1201         /*
1202          * Try to see if we can get the block without requesting a new
1203          * file system block.
1204          */
1205         down_read((&EXT4_I(inode)->i_data_sem));
1206         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1207                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1208                                 bh, 0);
1209         } else {
1210                 retval = ext4_ind_get_blocks(handle, inode, block, max_blocks,
1211                                              bh, 0);
1212         }
1213         up_read((&EXT4_I(inode)->i_data_sem));
1214
1215         if (retval > 0 && buffer_mapped(bh)) {
1216                 int ret = check_block_validity(inode, "file system corruption",
1217                                                block, bh->b_blocknr, retval);
1218                 if (ret != 0)
1219                         return ret;
1220         }
1221
1222         /* If it is only a block(s) look up */
1223         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
1224                 return retval;
1225
1226         /*
1227          * Returns if the blocks have already allocated
1228          *
1229          * Note that if blocks have been preallocated
1230          * ext4_ext_get_block() returns th create = 0
1231          * with buffer head unmapped.
1232          */
1233         if (retval > 0 && buffer_mapped(bh))
1234                 return retval;
1235
1236         /*
1237          * When we call get_blocks without the create flag, the
1238          * BH_Unwritten flag could have gotten set if the blocks
1239          * requested were part of a uninitialized extent.  We need to
1240          * clear this flag now that we are committed to convert all or
1241          * part of the uninitialized extent to be an initialized
1242          * extent.  This is because we need to avoid the combination
1243          * of BH_Unwritten and BH_Mapped flags being simultaneously
1244          * set on the buffer_head.
1245          */
1246         clear_buffer_unwritten(bh);
1247
1248         /*
1249          * New blocks allocate and/or writing to uninitialized extent
1250          * will possibly result in updating i_data, so we take
1251          * the write lock of i_data_sem, and call get_blocks()
1252          * with create == 1 flag.
1253          */
1254         down_write((&EXT4_I(inode)->i_data_sem));
1255
1256         /*
1257          * if the caller is from delayed allocation writeout path
1258          * we have already reserved fs blocks for allocation
1259          * let the underlying get_block() function know to
1260          * avoid double accounting
1261          */
1262         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1263                 EXT4_I(inode)->i_delalloc_reserved_flag = 1;
1264         /*
1265          * We need to check for EXT4 here because migrate
1266          * could have changed the inode type in between
1267          */
1268         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
1269                 retval =  ext4_ext_get_blocks(handle, inode, block, max_blocks,
1270                                               bh, flags);
1271         } else {
1272                 retval = ext4_ind_get_blocks(handle, inode, block,
1273                                              max_blocks, bh, flags);
1274
1275                 if (retval > 0 && buffer_new(bh)) {
1276                         /*
1277                          * We allocated new blocks which will result in
1278                          * i_data's format changing.  Force the migrate
1279                          * to fail by clearing migrate flags
1280                          */
1281                         EXT4_I(inode)->i_state &= ~EXT4_STATE_EXT_MIGRATE;
1282                 }
1283         }
1284
1285         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
1286                 EXT4_I(inode)->i_delalloc_reserved_flag = 0;
1287
1288         /*
1289          * Update reserved blocks/metadata blocks after successful
1290          * block allocation which had been deferred till now.
1291          */
1292         if ((retval > 0) && (flags & EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE))
1293                 ext4_da_update_reserve_space(inode, retval);
1294
1295         up_write((&EXT4_I(inode)->i_data_sem));
1296         if (retval > 0 && buffer_mapped(bh)) {
1297                 int ret = check_block_validity(inode, "file system "
1298                                                "corruption after allocation",
1299                                                block, bh->b_blocknr, retval);
1300                 if (ret != 0)
1301                         return ret;
1302         }
1303         return retval;
1304 }
1305
1306 /* Maximum number of blocks we map for direct IO at once. */
1307 #define DIO_MAX_BLOCKS 4096
1308
1309 int ext4_get_block(struct inode *inode, sector_t iblock,
1310                    struct buffer_head *bh_result, int create)
1311 {
1312         handle_t *handle = ext4_journal_current_handle();
1313         int ret = 0, started = 0;
1314         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
1315         int dio_credits;
1316
1317         if (create && !handle) {
1318                 /* Direct IO write... */
1319                 if (max_blocks > DIO_MAX_BLOCKS)
1320                         max_blocks = DIO_MAX_BLOCKS;
1321                 dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
1322                 handle = ext4_journal_start(inode, dio_credits);
1323                 if (IS_ERR(handle)) {
1324                         ret = PTR_ERR(handle);
1325                         goto out;
1326                 }
1327                 started = 1;
1328         }
1329
1330         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
1331                               create ? EXT4_GET_BLOCKS_CREATE : 0);
1332         if (ret > 0) {
1333                 bh_result->b_size = (ret << inode->i_blkbits);
1334                 ret = 0;
1335         }
1336         if (started)
1337                 ext4_journal_stop(handle);
1338 out:
1339         return ret;
1340 }
1341
1342 /*
1343  * `handle' can be NULL if create is zero
1344  */
1345 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
1346                                 ext4_lblk_t block, int create, int *errp)
1347 {
1348         struct buffer_head dummy;
1349         int fatal = 0, err;
1350         int flags = 0;
1351
1352         J_ASSERT(handle != NULL || create == 0);
1353
1354         dummy.b_state = 0;
1355         dummy.b_blocknr = -1000;
1356         buffer_trace_init(&dummy.b_history);
1357         if (create)
1358                 flags |= EXT4_GET_BLOCKS_CREATE;
1359         err = ext4_get_blocks(handle, inode, block, 1, &dummy, flags);
1360         /*
1361          * ext4_get_blocks() returns number of blocks mapped. 0 in
1362          * case of a HOLE.
1363          */
1364         if (err > 0) {
1365                 if (err > 1)
1366                         WARN_ON(1);
1367                 err = 0;
1368         }
1369         *errp = err;
1370         if (!err && buffer_mapped(&dummy)) {
1371                 struct buffer_head *bh;
1372                 bh = sb_getblk(inode->i_sb, dummy.b_blocknr);
1373                 if (!bh) {
1374                         *errp = -EIO;
1375                         goto err;
1376                 }
1377                 if (buffer_new(&dummy)) {
1378                         J_ASSERT(create != 0);
1379                         J_ASSERT(handle != NULL);
1380
1381                         /*
1382                          * Now that we do not always journal data, we should
1383                          * keep in mind whether this should always journal the
1384                          * new buffer as metadata.  For now, regular file
1385                          * writes use ext4_get_block instead, so it's not a
1386                          * problem.
1387                          */
1388                         lock_buffer(bh);
1389                         BUFFER_TRACE(bh, "call get_create_access");
1390                         fatal = ext4_journal_get_create_access(handle, bh);
1391                         if (!fatal && !buffer_uptodate(bh)) {
1392                                 memset(bh->b_data, 0, inode->i_sb->s_blocksize);
1393                                 set_buffer_uptodate(bh);
1394                         }
1395                         unlock_buffer(bh);
1396                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
1397                         err = ext4_handle_dirty_metadata(handle, inode, bh);
1398                         if (!fatal)
1399                                 fatal = err;
1400                 } else {
1401                         BUFFER_TRACE(bh, "not a new buffer");
1402                 }
1403                 if (fatal) {
1404                         *errp = fatal;
1405                         brelse(bh);
1406                         bh = NULL;
1407                 }
1408                 return bh;
1409         }
1410 err:
1411         return NULL;
1412 }
1413
1414 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
1415                                ext4_lblk_t block, int create, int *err)
1416 {
1417         struct buffer_head *bh;
1418
1419         bh = ext4_getblk(handle, inode, block, create, err);
1420         if (!bh)
1421                 return bh;
1422         if (buffer_uptodate(bh))
1423                 return bh;
1424         ll_rw_block(READ_META, 1, &bh);
1425         wait_on_buffer(bh);
1426         if (buffer_uptodate(bh))
1427                 return bh;
1428         put_bh(bh);
1429         *err = -EIO;
1430         return NULL;
1431 }
1432
1433 static int walk_page_buffers(handle_t *handle,
1434                              struct buffer_head *head,
1435                              unsigned from,
1436                              unsigned to,
1437                              int *partial,
1438                              int (*fn)(handle_t *handle,
1439                                        struct buffer_head *bh))
1440 {
1441         struct buffer_head *bh;
1442         unsigned block_start, block_end;
1443         unsigned blocksize = head->b_size;
1444         int err, ret = 0;
1445         struct buffer_head *next;
1446
1447         for (bh = head, block_start = 0;
1448              ret == 0 && (bh != head || !block_start);
1449              block_start = block_end, bh = next) {
1450                 next = bh->b_this_page;
1451                 block_end = block_start + blocksize;
1452                 if (block_end <= from || block_start >= to) {
1453                         if (partial && !buffer_uptodate(bh))
1454                                 *partial = 1;
1455                         continue;
1456                 }
1457                 err = (*fn)(handle, bh);
1458                 if (!ret)
1459                         ret = err;
1460         }
1461         return ret;
1462 }
1463
1464 /*
1465  * To preserve ordering, it is essential that the hole instantiation and
1466  * the data write be encapsulated in a single transaction.  We cannot
1467  * close off a transaction and start a new one between the ext4_get_block()
1468  * and the commit_write().  So doing the jbd2_journal_start at the start of
1469  * prepare_write() is the right place.
1470  *
1471  * Also, this function can nest inside ext4_writepage() ->
1472  * block_write_full_page(). In that case, we *know* that ext4_writepage()
1473  * has generated enough buffer credits to do the whole page.  So we won't
1474  * block on the journal in that case, which is good, because the caller may
1475  * be PF_MEMALLOC.
1476  *
1477  * By accident, ext4 can be reentered when a transaction is open via
1478  * quota file writes.  If we were to commit the transaction while thus
1479  * reentered, there can be a deadlock - we would be holding a quota
1480  * lock, and the commit would never complete if another thread had a
1481  * transaction open and was blocking on the quota lock - a ranking
1482  * violation.
1483  *
1484  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
1485  * will _not_ run commit under these circumstances because handle->h_ref
1486  * is elevated.  We'll still have enough credits for the tiny quotafile
1487  * write.
1488  */
1489 static int do_journal_get_write_access(handle_t *handle,
1490                                        struct buffer_head *bh)
1491 {
1492         if (!buffer_mapped(bh) || buffer_freed(bh))
1493                 return 0;
1494         return ext4_journal_get_write_access(handle, bh);
1495 }
1496
1497 /*
1498  * Truncate blocks that were not used by write. We have to truncate the
1499  * pagecache as well so that corresponding buffers get properly unmapped.
1500  */
1501 static void ext4_truncate_failed_write(struct inode *inode)
1502 {
1503         truncate_inode_pages(inode->i_mapping, inode->i_size);
1504         ext4_truncate(inode);
1505 }
1506
1507 static int ext4_write_begin(struct file *file, struct address_space *mapping,
1508                             loff_t pos, unsigned len, unsigned flags,
1509                             struct page **pagep, void **fsdata)
1510 {
1511         struct inode *inode = mapping->host;
1512         int ret, needed_blocks;
1513         handle_t *handle;
1514         int retries = 0;
1515         struct page *page;
1516         pgoff_t index;
1517         unsigned from, to;
1518
1519         trace_ext4_write_begin(inode, pos, len, flags);
1520         /*
1521          * Reserve one block more for addition to orphan list in case
1522          * we allocate blocks but write fails for some reason
1523          */
1524         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
1525         index = pos >> PAGE_CACHE_SHIFT;
1526         from = pos & (PAGE_CACHE_SIZE - 1);
1527         to = from + len;
1528
1529 retry:
1530         handle = ext4_journal_start(inode, needed_blocks);
1531         if (IS_ERR(handle)) {
1532                 ret = PTR_ERR(handle);
1533                 goto out;
1534         }
1535
1536         /* We cannot recurse into the filesystem as the transaction is already
1537          * started */
1538         flags |= AOP_FLAG_NOFS;
1539
1540         page = grab_cache_page_write_begin(mapping, index, flags);
1541         if (!page) {
1542                 ext4_journal_stop(handle);
1543                 ret = -ENOMEM;
1544                 goto out;
1545         }
1546         *pagep = page;
1547
1548         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1549                                 ext4_get_block);
1550
1551         if (!ret && ext4_should_journal_data(inode)) {
1552                 ret = walk_page_buffers(handle, page_buffers(page),
1553                                 from, to, NULL, do_journal_get_write_access);
1554         }
1555
1556         if (ret) {
1557                 unlock_page(page);
1558                 page_cache_release(page);
1559                 /*
1560                  * block_write_begin may have instantiated a few blocks
1561                  * outside i_size.  Trim these off again. Don't need
1562                  * i_size_read because we hold i_mutex.
1563                  *
1564                  * Add inode to orphan list in case we crash before
1565                  * truncate finishes
1566                  */
1567                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1568                         ext4_orphan_add(handle, inode);
1569
1570                 ext4_journal_stop(handle);
1571                 if (pos + len > inode->i_size) {
1572                         ext4_truncate_failed_write(inode);
1573                         /*
1574                          * If truncate failed early the inode might
1575                          * still be on the orphan list; we need to
1576                          * make sure the inode is removed from the
1577                          * orphan list in that case.
1578                          */
1579                         if (inode->i_nlink)
1580                                 ext4_orphan_del(NULL, inode);
1581                 }
1582         }
1583
1584         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
1585                 goto retry;
1586 out:
1587         return ret;
1588 }
1589
1590 /* For write_end() in data=journal mode */
1591 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
1592 {
1593         if (!buffer_mapped(bh) || buffer_freed(bh))
1594                 return 0;
1595         set_buffer_uptodate(bh);
1596         return ext4_handle_dirty_metadata(handle, NULL, bh);
1597 }
1598
1599 static int ext4_generic_write_end(struct file *file,
1600                                   struct address_space *mapping,
1601                                   loff_t pos, unsigned len, unsigned copied,
1602                                   struct page *page, void *fsdata)
1603 {
1604         int i_size_changed = 0;
1605         struct inode *inode = mapping->host;
1606         handle_t *handle = ext4_journal_current_handle();
1607
1608         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
1609
1610         /*
1611          * No need to use i_size_read() here, the i_size
1612          * cannot change under us because we hold i_mutex.
1613          *
1614          * But it's important to update i_size while still holding page lock:
1615          * page writeout could otherwise come in and zero beyond i_size.
1616          */
1617         if (pos + copied > inode->i_size) {
1618                 i_size_write(inode, pos + copied);
1619                 i_size_changed = 1;
1620         }
1621
1622         if (pos + copied >  EXT4_I(inode)->i_disksize) {
1623                 /* We need to mark inode dirty even if
1624                  * new_i_size is less that inode->i_size
1625                  * bu greater than i_disksize.(hint delalloc)
1626                  */
1627                 ext4_update_i_disksize(inode, (pos + copied));
1628                 i_size_changed = 1;
1629         }
1630         unlock_page(page);
1631         page_cache_release(page);
1632
1633         /*
1634          * Don't mark the inode dirty under page lock. First, it unnecessarily
1635          * makes the holding time of page lock longer. Second, it forces lock
1636          * ordering of page lock and transaction start for journaling
1637          * filesystems.
1638          */
1639         if (i_size_changed)
1640                 ext4_mark_inode_dirty(handle, inode);
1641
1642         return copied;
1643 }
1644
1645 /*
1646  * We need to pick up the new inode size which generic_commit_write gave us
1647  * `file' can be NULL - eg, when called from page_symlink().
1648  *
1649  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
1650  * buffers are managed internally.
1651  */
1652 static int ext4_ordered_write_end(struct file *file,
1653                                   struct address_space *mapping,
1654                                   loff_t pos, unsigned len, unsigned copied,
1655                                   struct page *page, void *fsdata)
1656 {
1657         handle_t *handle = ext4_journal_current_handle();
1658         struct inode *inode = mapping->host;
1659         int ret = 0, ret2;
1660
1661         trace_ext4_ordered_write_end(inode, pos, len, copied);
1662         ret = ext4_jbd2_file_inode(handle, inode);
1663
1664         if (ret == 0) {
1665                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1666                                                         page, fsdata);
1667                 copied = ret2;
1668                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
1669                         /* if we have allocated more blocks and copied
1670                          * less. We will have blocks allocated outside
1671                          * inode->i_size. So truncate them
1672                          */
1673                         ext4_orphan_add(handle, inode);
1674                 if (ret2 < 0)
1675                         ret = ret2;
1676         }
1677         ret2 = ext4_journal_stop(handle);
1678         if (!ret)
1679                 ret = ret2;
1680
1681         if (pos + len > inode->i_size) {
1682                 ext4_truncate_failed_write(inode);
1683                 /*
1684                  * If truncate failed early the inode might still be
1685                  * on the orphan list; we need to make sure the inode
1686                  * is removed from the orphan list in that case.
1687                  */
1688                 if (inode->i_nlink)
1689                         ext4_orphan_del(NULL, inode);
1690         }
1691
1692
1693         return ret ? ret : copied;
1694 }
1695
1696 static int ext4_writeback_write_end(struct file *file,
1697                                     struct address_space *mapping,
1698                                     loff_t pos, unsigned len, unsigned copied,
1699                                     struct page *page, void *fsdata)
1700 {
1701         handle_t *handle = ext4_journal_current_handle();
1702         struct inode *inode = mapping->host;
1703         int ret = 0, ret2;
1704
1705         trace_ext4_writeback_write_end(inode, pos, len, copied);
1706         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1707                                                         page, fsdata);
1708         copied = ret2;
1709         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1710                 /* if we have allocated more blocks and copied
1711                  * less. We will have blocks allocated outside
1712                  * inode->i_size. So truncate them
1713                  */
1714                 ext4_orphan_add(handle, inode);
1715
1716         if (ret2 < 0)
1717                 ret = ret2;
1718
1719         ret2 = ext4_journal_stop(handle);
1720         if (!ret)
1721                 ret = ret2;
1722
1723         if (pos + len > inode->i_size) {
1724                 ext4_truncate_failed_write(inode);
1725                 /*
1726                  * If truncate failed early the inode might still be
1727                  * on the orphan list; we need to make sure the inode
1728                  * is removed from the orphan list in that case.
1729                  */
1730                 if (inode->i_nlink)
1731                         ext4_orphan_del(NULL, inode);
1732         }
1733
1734         return ret ? ret : copied;
1735 }
1736
1737 static int ext4_journalled_write_end(struct file *file,
1738                                      struct address_space *mapping,
1739                                      loff_t pos, unsigned len, unsigned copied,
1740                                      struct page *page, void *fsdata)
1741 {
1742         handle_t *handle = ext4_journal_current_handle();
1743         struct inode *inode = mapping->host;
1744         int ret = 0, ret2;
1745         int partial = 0;
1746         unsigned from, to;
1747         loff_t new_i_size;
1748
1749         trace_ext4_journalled_write_end(inode, pos, len, copied);
1750         from = pos & (PAGE_CACHE_SIZE - 1);
1751         to = from + len;
1752
1753         if (copied < len) {
1754                 if (!PageUptodate(page))
1755                         copied = 0;
1756                 page_zero_new_buffers(page, from+copied, to);
1757         }
1758
1759         ret = walk_page_buffers(handle, page_buffers(page), from,
1760                                 to, &partial, write_end_fn);
1761         if (!partial)
1762                 SetPageUptodate(page);
1763         new_i_size = pos + copied;
1764         if (new_i_size > inode->i_size)
1765                 i_size_write(inode, pos+copied);
1766         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
1767         if (new_i_size > EXT4_I(inode)->i_disksize) {
1768                 ext4_update_i_disksize(inode, new_i_size);
1769                 ret2 = ext4_mark_inode_dirty(handle, inode);
1770                 if (!ret)
1771                         ret = ret2;
1772         }
1773
1774         unlock_page(page);
1775         page_cache_release(page);
1776         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1777                 /* if we have allocated more blocks and copied
1778                  * less. We will have blocks allocated outside
1779                  * inode->i_size. So truncate them
1780                  */
1781                 ext4_orphan_add(handle, inode);
1782
1783         ret2 = ext4_journal_stop(handle);
1784         if (!ret)
1785                 ret = ret2;
1786         if (pos + len > inode->i_size) {
1787                 ext4_truncate_failed_write(inode);
1788                 /*
1789                  * If truncate failed early the inode might still be
1790                  * on the orphan list; we need to make sure the inode
1791                  * is removed from the orphan list in that case.
1792                  */
1793                 if (inode->i_nlink)
1794                         ext4_orphan_del(NULL, inode);
1795         }
1796
1797         return ret ? ret : copied;
1798 }
1799
1800 static int ext4_da_reserve_space(struct inode *inode, int nrblocks)
1801 {
1802         int retries = 0;
1803         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1804         unsigned long md_needed, mdblocks, total = 0;
1805
1806         /*
1807          * recalculate the amount of metadata blocks to reserve
1808          * in order to allocate nrblocks
1809          * worse case is one extent per block
1810          */
1811 repeat:
1812         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1813         total = EXT4_I(inode)->i_reserved_data_blocks + nrblocks;
1814         mdblocks = ext4_calc_metadata_amount(inode, total);
1815         BUG_ON(mdblocks < EXT4_I(inode)->i_reserved_meta_blocks);
1816
1817         md_needed = mdblocks - EXT4_I(inode)->i_reserved_meta_blocks;
1818         total = md_needed + nrblocks;
1819         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1820
1821         /*
1822          * Make quota reservation here to prevent quota overflow
1823          * later. Real quota accounting is done at pages writeout
1824          * time.
1825          */
1826         if (vfs_dq_reserve_block(inode, total))
1827                 return -EDQUOT;
1828
1829         if (ext4_claim_free_blocks(sbi, total)) {
1830                 vfs_dq_release_reservation_block(inode, total);
1831                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1832                         yield();
1833                         goto repeat;
1834                 }
1835                 return -ENOSPC;
1836         }
1837         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1838         EXT4_I(inode)->i_reserved_data_blocks += nrblocks;
1839         EXT4_I(inode)->i_reserved_meta_blocks += md_needed;
1840         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1841
1842         return 0;       /* success */
1843 }
1844
1845 static void ext4_da_release_space(struct inode *inode, int to_free)
1846 {
1847         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1848         int total, mdb, mdb_free, release;
1849
1850         if (!to_free)
1851                 return;         /* Nothing to release, exit */
1852
1853         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1854
1855         if (!EXT4_I(inode)->i_reserved_data_blocks) {
1856                 /*
1857                  * if there is no reserved blocks, but we try to free some
1858                  * then the counter is messed up somewhere.
1859                  * but since this function is called from invalidate
1860                  * page, it's harmless to return without any action
1861                  */
1862                 printk(KERN_INFO "ext4 delalloc try to release %d reserved "
1863                             "blocks for inode %lu, but there is no reserved "
1864                             "data blocks\n", to_free, inode->i_ino);
1865                 spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1866                 return;
1867         }
1868
1869         /* recalculate the number of metablocks still need to be reserved */
1870         total = EXT4_I(inode)->i_reserved_data_blocks - to_free;
1871         mdb = ext4_calc_metadata_amount(inode, total);
1872
1873         /* figure out how many metablocks to release */
1874         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1875         mdb_free = EXT4_I(inode)->i_reserved_meta_blocks - mdb;
1876
1877         release = to_free + mdb_free;
1878
1879         /* update fs dirty blocks counter for truncate case */
1880         percpu_counter_sub(&sbi->s_dirtyblocks_counter, release);
1881
1882         /* update per-inode reservations */
1883         BUG_ON(to_free > EXT4_I(inode)->i_reserved_data_blocks);
1884         EXT4_I(inode)->i_reserved_data_blocks -= to_free;
1885
1886         BUG_ON(mdb > EXT4_I(inode)->i_reserved_meta_blocks);
1887         EXT4_I(inode)->i_reserved_meta_blocks = mdb;
1888         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1889
1890         vfs_dq_release_reservation_block(inode, release);
1891 }
1892
1893 static void ext4_da_page_release_reservation(struct page *page,
1894                                              unsigned long offset)
1895 {
1896         int to_release = 0;
1897         struct buffer_head *head, *bh;
1898         unsigned int curr_off = 0;
1899
1900         head = page_buffers(page);
1901         bh = head;
1902         do {
1903                 unsigned int next_off = curr_off + bh->b_size;
1904
1905                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1906                         to_release++;
1907                         clear_buffer_delay(bh);
1908                 }
1909                 curr_off = next_off;
1910         } while ((bh = bh->b_this_page) != head);
1911         ext4_da_release_space(page->mapping->host, to_release);
1912 }
1913
1914 /*
1915  * Delayed allocation stuff
1916  */
1917
1918 /*
1919  * mpage_da_submit_io - walks through extent of pages and try to write
1920  * them with writepage() call back
1921  *
1922  * @mpd->inode: inode
1923  * @mpd->first_page: first page of the extent
1924  * @mpd->next_page: page after the last page of the extent
1925  *
1926  * By the time mpage_da_submit_io() is called we expect all blocks
1927  * to be allocated. this may be wrong if allocation failed.
1928  *
1929  * As pages are already locked by write_cache_pages(), we can't use it
1930  */
1931 static int mpage_da_submit_io(struct mpage_da_data *mpd)
1932 {
1933         long pages_skipped;
1934         struct pagevec pvec;
1935         unsigned long index, end;
1936         int ret = 0, err, nr_pages, i;
1937         struct inode *inode = mpd->inode;
1938         struct address_space *mapping = inode->i_mapping;
1939
1940         BUG_ON(mpd->next_page <= mpd->first_page);
1941         /*
1942          * We need to start from the first_page to the next_page - 1
1943          * to make sure we also write the mapped dirty buffer_heads.
1944          * If we look at mpd->b_blocknr we would only be looking
1945          * at the currently mapped buffer_heads.
1946          */
1947         index = mpd->first_page;
1948         end = mpd->next_page - 1;
1949
1950         pagevec_init(&pvec, 0);
1951         while (index <= end) {
1952                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1953                 if (nr_pages == 0)
1954                         break;
1955                 for (i = 0; i < nr_pages; i++) {
1956                         struct page *page = pvec.pages[i];
1957
1958                         index = page->index;
1959                         if (index > end)
1960                                 break;
1961                         index++;
1962
1963                         BUG_ON(!PageLocked(page));
1964                         BUG_ON(PageWriteback(page));
1965
1966                         pages_skipped = mpd->wbc->pages_skipped;
1967                         err = mapping->a_ops->writepage(page, mpd->wbc);
1968                         if (!err && (pages_skipped == mpd->wbc->pages_skipped))
1969                                 /*
1970                                  * have successfully written the page
1971                                  * without skipping the same
1972                                  */
1973                                 mpd->pages_written++;
1974                         /*
1975                          * In error case, we have to continue because
1976                          * remaining pages are still locked
1977                          * XXX: unlock and re-dirty them?
1978                          */
1979                         if (ret == 0)
1980                                 ret = err;
1981                 }
1982                 pagevec_release(&pvec);
1983         }
1984         return ret;
1985 }
1986
1987 /*
1988  * mpage_put_bnr_to_bhs - walk blocks and assign them actual numbers
1989  *
1990  * @mpd->inode - inode to walk through
1991  * @exbh->b_blocknr - first block on a disk
1992  * @exbh->b_size - amount of space in bytes
1993  * @logical - first logical block to start assignment with
1994  *
1995  * the function goes through all passed space and put actual disk
1996  * block numbers into buffer heads, dropping BH_Delay and BH_Unwritten
1997  */
1998 static void mpage_put_bnr_to_bhs(struct mpage_da_data *mpd, sector_t logical,
1999                                  struct buffer_head *exbh)
2000 {
2001         struct inode *inode = mpd->inode;
2002         struct address_space *mapping = inode->i_mapping;
2003         int blocks = exbh->b_size >> inode->i_blkbits;
2004         sector_t pblock = exbh->b_blocknr, cur_logical;
2005         struct buffer_head *head, *bh;
2006         pgoff_t index, end;
2007         struct pagevec pvec;
2008         int nr_pages, i;
2009
2010         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2011         end = (logical + blocks - 1) >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2012         cur_logical = index << (PAGE_CACHE_SHIFT - inode->i_blkbits);
2013
2014         pagevec_init(&pvec, 0);
2015
2016         while (index <= end) {
2017                 /* XXX: optimize tail */
2018                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2019                 if (nr_pages == 0)
2020                         break;
2021                 for (i = 0; i < nr_pages; i++) {
2022                         struct page *page = pvec.pages[i];
2023
2024                         index = page->index;
2025                         if (index > end)
2026                                 break;
2027                         index++;
2028
2029                         BUG_ON(!PageLocked(page));
2030                         BUG_ON(PageWriteback(page));
2031                         BUG_ON(!page_has_buffers(page));
2032
2033                         bh = page_buffers(page);
2034                         head = bh;
2035
2036                         /* skip blocks out of the range */
2037                         do {
2038                                 if (cur_logical >= logical)
2039                                         break;
2040                                 cur_logical++;
2041                         } while ((bh = bh->b_this_page) != head);
2042
2043                         do {
2044                                 if (cur_logical >= logical + blocks)
2045                                         break;
2046
2047                                 if (buffer_delay(bh) ||
2048                                                 buffer_unwritten(bh)) {
2049
2050                                         BUG_ON(bh->b_bdev != inode->i_sb->s_bdev);
2051
2052                                         if (buffer_delay(bh)) {
2053                                                 clear_buffer_delay(bh);
2054                                                 bh->b_blocknr = pblock;
2055                                         } else {
2056                                                 /*
2057                                                  * unwritten already should have
2058                                                  * blocknr assigned. Verify that
2059                                                  */
2060                                                 clear_buffer_unwritten(bh);
2061                                                 BUG_ON(bh->b_blocknr != pblock);
2062                                         }
2063
2064                                 } else if (buffer_mapped(bh))
2065                                         BUG_ON(bh->b_blocknr != pblock);
2066
2067                                 cur_logical++;
2068                                 pblock++;
2069                         } while ((bh = bh->b_this_page) != head);
2070                 }
2071                 pagevec_release(&pvec);
2072         }
2073 }
2074
2075
2076 /*
2077  * __unmap_underlying_blocks - just a helper function to unmap
2078  * set of blocks described by @bh
2079  */
2080 static inline void __unmap_underlying_blocks(struct inode *inode,
2081                                              struct buffer_head *bh)
2082 {
2083         struct block_device *bdev = inode->i_sb->s_bdev;
2084         int blocks, i;
2085
2086         blocks = bh->b_size >> inode->i_blkbits;
2087         for (i = 0; i < blocks; i++)
2088                 unmap_underlying_metadata(bdev, bh->b_blocknr + i);
2089 }
2090
2091 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd,
2092                                         sector_t logical, long blk_cnt)
2093 {
2094         int nr_pages, i;
2095         pgoff_t index, end;
2096         struct pagevec pvec;
2097         struct inode *inode = mpd->inode;
2098         struct address_space *mapping = inode->i_mapping;
2099
2100         index = logical >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
2101         end   = (logical + blk_cnt - 1) >>
2102                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2103         while (index <= end) {
2104                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
2105                 if (nr_pages == 0)
2106                         break;
2107                 for (i = 0; i < nr_pages; i++) {
2108                         struct page *page = pvec.pages[i];
2109                         index = page->index;
2110                         if (index > end)
2111                                 break;
2112                         index++;
2113
2114                         BUG_ON(!PageLocked(page));
2115                         BUG_ON(PageWriteback(page));
2116                         block_invalidatepage(page, 0);
2117                         ClearPageUptodate(page);
2118                         unlock_page(page);
2119                 }
2120         }
2121         return;
2122 }
2123
2124 static void ext4_print_free_blocks(struct inode *inode)
2125 {
2126         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2127         printk(KERN_CRIT "Total free blocks count %lld\n",
2128                ext4_count_free_blocks(inode->i_sb));
2129         printk(KERN_CRIT "Free/Dirty block details\n");
2130         printk(KERN_CRIT "free_blocks=%lld\n",
2131                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
2132         printk(KERN_CRIT "dirty_blocks=%lld\n",
2133                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
2134         printk(KERN_CRIT "Block reservation details\n");
2135         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
2136                EXT4_I(inode)->i_reserved_data_blocks);
2137         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
2138                EXT4_I(inode)->i_reserved_meta_blocks);
2139         return;
2140 }
2141
2142 /*
2143  * mpage_da_map_blocks - go through given space
2144  *
2145  * @mpd - bh describing space
2146  *
2147  * The function skips space we know is already mapped to disk blocks.
2148  *
2149  */
2150 static int mpage_da_map_blocks(struct mpage_da_data *mpd)
2151 {
2152         int err, blks, get_blocks_flags;
2153         struct buffer_head new;
2154         sector_t next = mpd->b_blocknr;
2155         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
2156         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
2157         handle_t *handle = NULL;
2158
2159         /*
2160          * We consider only non-mapped and non-allocated blocks
2161          */
2162         if ((mpd->b_state  & (1 << BH_Mapped)) &&
2163                 !(mpd->b_state & (1 << BH_Delay)) &&
2164                 !(mpd->b_state & (1 << BH_Unwritten)))
2165                 return 0;
2166
2167         /*
2168          * If we didn't accumulate anything to write simply return
2169          */
2170         if (!mpd->b_size)
2171                 return 0;
2172
2173         handle = ext4_journal_current_handle();
2174         BUG_ON(!handle);
2175
2176         /*
2177          * Call ext4_get_blocks() to allocate any delayed allocation
2178          * blocks, or to convert an uninitialized extent to be
2179          * initialized (in the case where we have written into
2180          * one or more preallocated blocks).
2181          *
2182          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
2183          * indicate that we are on the delayed allocation path.  This
2184          * affects functions in many different parts of the allocation
2185          * call path.  This flag exists primarily because we don't
2186          * want to change *many* call functions, so ext4_get_blocks()
2187          * will set the magic i_delalloc_reserved_flag once the
2188          * inode's allocation semaphore is taken.
2189          *
2190          * If the blocks in questions were delalloc blocks, set
2191          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
2192          * variables are updated after the blocks have been allocated.
2193          */
2194         new.b_state = 0;
2195         get_blocks_flags = (EXT4_GET_BLOCKS_CREATE |
2196                             EXT4_GET_BLOCKS_DELALLOC_RESERVE);
2197         if (mpd->b_state & (1 << BH_Delay))
2198                 get_blocks_flags |= EXT4_GET_BLOCKS_UPDATE_RESERVE_SPACE;
2199         blks = ext4_get_blocks(handle, mpd->inode, next, max_blocks,
2200                                &new, get_blocks_flags);
2201         if (blks < 0) {
2202                 err = blks;
2203                 /*
2204                  * If get block returns with error we simply
2205                  * return. Later writepage will redirty the page and
2206                  * writepages will find the dirty page again
2207                  */
2208                 if (err == -EAGAIN)
2209                         return 0;
2210
2211                 if (err == -ENOSPC &&
2212                     ext4_count_free_blocks(mpd->inode->i_sb)) {
2213                         mpd->retval = err;
2214                         return 0;
2215                 }
2216
2217                 /*
2218                  * get block failure will cause us to loop in
2219                  * writepages, because a_ops->writepage won't be able
2220                  * to make progress. The page will be redirtied by
2221                  * writepage and writepages will again try to write
2222                  * the same.
2223                  */
2224                 ext4_msg(mpd->inode->i_sb, KERN_CRIT,
2225                          "delayed block allocation failed for inode %lu at "
2226                          "logical offset %llu with max blocks %zd with "
2227                          "error %d\n", mpd->inode->i_ino,
2228                          (unsigned long long) next,
2229                          mpd->b_size >> mpd->inode->i_blkbits, err);
2230                 printk(KERN_CRIT "This should not happen!!  "
2231                        "Data will be lost\n");
2232                 if (err == -ENOSPC) {
2233                         ext4_print_free_blocks(mpd->inode);
2234                 }
2235                 /* invalidate all the pages */
2236                 ext4_da_block_invalidatepages(mpd, next,
2237                                 mpd->b_size >> mpd->inode->i_blkbits);
2238                 return err;
2239         }
2240         BUG_ON(blks == 0);
2241
2242         new.b_size = (blks << mpd->inode->i_blkbits);
2243
2244         if (buffer_new(&new))
2245                 __unmap_underlying_blocks(mpd->inode, &new);
2246
2247         /*
2248          * If blocks are delayed marked, we need to
2249          * put actual blocknr and drop delayed bit
2250          */
2251         if ((mpd->b_state & (1 << BH_Delay)) ||
2252             (mpd->b_state & (1 << BH_Unwritten)))
2253                 mpage_put_bnr_to_bhs(mpd, next, &new);
2254
2255         if (ext4_should_order_data(mpd->inode)) {
2256                 err = ext4_jbd2_file_inode(handle, mpd->inode);
2257                 if (err)
2258                         return err;
2259         }
2260
2261         /*
2262          * Update on-disk size along with block allocation.
2263          */
2264         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
2265         if (disksize > i_size_read(mpd->inode))
2266                 disksize = i_size_read(mpd->inode);
2267         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
2268                 ext4_update_i_disksize(mpd->inode, disksize);
2269                 return ext4_mark_inode_dirty(handle, mpd->inode);
2270         }
2271
2272         return 0;
2273 }
2274
2275 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
2276                 (1 << BH_Delay) | (1 << BH_Unwritten))
2277
2278 /*
2279  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
2280  *
2281  * @mpd->lbh - extent of blocks
2282  * @logical - logical number of the block in the file
2283  * @bh - bh of the block (used to access block's state)
2284  *
2285  * the function is used to collect contig. blocks in same state
2286  */
2287 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
2288                                    sector_t logical, size_t b_size,
2289                                    unsigned long b_state)
2290 {
2291         sector_t next;
2292         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
2293
2294         /* check if thereserved journal credits might overflow */
2295         if (!(EXT4_I(mpd->inode)->i_flags & EXT4_EXTENTS_FL)) {
2296                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
2297                         /*
2298                          * With non-extent format we are limited by the journal
2299                          * credit available.  Total credit needed to insert
2300                          * nrblocks contiguous blocks is dependent on the
2301                          * nrblocks.  So limit nrblocks.
2302                          */
2303                         goto flush_it;
2304                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
2305                                 EXT4_MAX_TRANS_DATA) {
2306                         /*
2307                          * Adding the new buffer_head would make it cross the
2308                          * allowed limit for which we have journal credit
2309                          * reserved. So limit the new bh->b_size
2310                          */
2311                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
2312                                                 mpd->inode->i_blkbits;
2313                         /* we will do mpage_da_submit_io in the next loop */
2314                 }
2315         }
2316         /*
2317          * First block in the extent
2318          */
2319         if (mpd->b_size == 0) {
2320                 mpd->b_blocknr = logical;
2321                 mpd->b_size = b_size;
2322                 mpd->b_state = b_state & BH_FLAGS;
2323                 return;
2324         }
2325
2326         next = mpd->b_blocknr + nrblocks;
2327         /*
2328          * Can we merge the block to our big extent?
2329          */
2330         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
2331                 mpd->b_size += b_size;
2332                 return;
2333         }
2334
2335 flush_it:
2336         /*
2337          * We couldn't merge the block to our extent, so we
2338          * need to flush current  extent and start new one
2339          */
2340         if (mpage_da_map_blocks(mpd) == 0)
2341                 mpage_da_submit_io(mpd);
2342         mpd->io_done = 1;
2343         return;
2344 }
2345
2346 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
2347 {
2348         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
2349 }
2350
2351 /*
2352  * __mpage_da_writepage - finds extent of pages and blocks
2353  *
2354  * @page: page to consider
2355  * @wbc: not used, we just follow rules
2356  * @data: context
2357  *
2358  * The function finds extents of pages and scan them for all blocks.
2359  */
2360 static int __mpage_da_writepage(struct page *page,
2361                                 struct writeback_control *wbc, void *data)
2362 {
2363         struct mpage_da_data *mpd = data;
2364         struct inode *inode = mpd->inode;
2365         struct buffer_head *bh, *head;
2366         sector_t logical;
2367
2368         if (mpd->io_done) {
2369                 /*
2370                  * Rest of the page in the page_vec
2371                  * redirty then and skip then. We will
2372                  * try to write them again after
2373                  * starting a new transaction
2374                  */
2375                 redirty_page_for_writepage(wbc, page);
2376                 unlock_page(page);
2377                 return MPAGE_DA_EXTENT_TAIL;
2378         }
2379         /*
2380          * Can we merge this page to current extent?
2381          */
2382         if (mpd->next_page != page->index) {
2383                 /*
2384                  * Nope, we can't. So, we map non-allocated blocks
2385                  * and start IO on them using writepage()
2386                  */
2387                 if (mpd->next_page != mpd->first_page) {
2388                         if (mpage_da_map_blocks(mpd) == 0)
2389                                 mpage_da_submit_io(mpd);
2390                         /*
2391                          * skip rest of the page in the page_vec
2392                          */
2393                         mpd->io_done = 1;
2394                         redirty_page_for_writepage(wbc, page);
2395                         unlock_page(page);
2396                         return MPAGE_DA_EXTENT_TAIL;
2397                 }
2398
2399                 /*
2400                  * Start next extent of pages ...
2401                  */
2402                 mpd->first_page = page->index;
2403
2404                 /*
2405                  * ... and blocks
2406                  */
2407                 mpd->b_size = 0;
2408                 mpd->b_state = 0;
2409                 mpd->b_blocknr = 0;
2410         }
2411
2412         mpd->next_page = page->index + 1;
2413         logical = (sector_t) page->index <<
2414                   (PAGE_CACHE_SHIFT - inode->i_blkbits);
2415
2416         if (!page_has_buffers(page)) {
2417                 mpage_add_bh_to_extent(mpd, logical, PAGE_CACHE_SIZE,
2418                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2419                 if (mpd->io_done)
2420                         return MPAGE_DA_EXTENT_TAIL;
2421         } else {
2422                 /*
2423                  * Page with regular buffer heads, just add all dirty ones
2424                  */
2425                 head = page_buffers(page);
2426                 bh = head;
2427                 do {
2428                         BUG_ON(buffer_locked(bh));
2429                         /*
2430                          * We need to try to allocate
2431                          * unmapped blocks in the same page.
2432                          * Otherwise we won't make progress
2433                          * with the page in ext4_writepage
2434                          */
2435                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2436                                 mpage_add_bh_to_extent(mpd, logical,
2437                                                        bh->b_size,
2438                                                        bh->b_state);
2439                                 if (mpd->io_done)
2440                                         return MPAGE_DA_EXTENT_TAIL;
2441                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2442                                 /*
2443                                  * mapped dirty buffer. We need to update
2444                                  * the b_state because we look at
2445                                  * b_state in mpage_da_map_blocks. We don't
2446                                  * update b_size because if we find an
2447                                  * unmapped buffer_head later we need to
2448                                  * use the b_state flag of that buffer_head.
2449                                  */
2450                                 if (mpd->b_size == 0)
2451                                         mpd->b_state = bh->b_state & BH_FLAGS;
2452                         }
2453                         logical++;
2454                 } while ((bh = bh->b_this_page) != head);
2455         }
2456
2457         return 0;
2458 }
2459
2460 /*
2461  * This is a special get_blocks_t callback which is used by
2462  * ext4_da_write_begin().  It will either return mapped block or
2463  * reserve space for a single block.
2464  *
2465  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
2466  * We also have b_blocknr = -1 and b_bdev initialized properly
2467  *
2468  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
2469  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
2470  * initialized properly.
2471  */
2472 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
2473                                   struct buffer_head *bh_result, int create)
2474 {
2475         int ret = 0;
2476         sector_t invalid_block = ~((sector_t) 0xffff);
2477
2478         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
2479                 invalid_block = ~0;
2480
2481         BUG_ON(create == 0);
2482         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2483
2484         /*
2485          * first, we need to know whether the block is allocated already
2486          * preallocated blocks are unmapped but should treated
2487          * the same as allocated blocks.
2488          */
2489         ret = ext4_get_blocks(NULL, inode, iblock, 1,  bh_result, 0);
2490         if ((ret == 0) && !buffer_delay(bh_result)) {
2491                 /* the block isn't (pre)allocated yet, let's reserve space */
2492                 /*
2493                  * XXX: __block_prepare_write() unmaps passed block,
2494                  * is it OK?
2495                  */
2496                 ret = ext4_da_reserve_space(inode, 1);
2497                 if (ret)
2498                         /* not enough space to reserve */
2499                         return ret;
2500
2501                 map_bh(bh_result, inode->i_sb, invalid_block);
2502                 set_buffer_new(bh_result);
2503                 set_buffer_delay(bh_result);
2504         } else if (ret > 0) {
2505                 bh_result->b_size = (ret << inode->i_blkbits);
2506                 if (buffer_unwritten(bh_result)) {
2507                         /* A delayed write to unwritten bh should
2508                          * be marked new and mapped.  Mapped ensures
2509                          * that we don't do get_block multiple times
2510                          * when we write to the same offset and new
2511                          * ensures that we do proper zero out for
2512                          * partial write.
2513                          */
2514                         set_buffer_new(bh_result);
2515                         set_buffer_mapped(bh_result);
2516                 }
2517                 ret = 0;
2518         }
2519
2520         return ret;
2521 }
2522
2523 /*
2524  * This function is used as a standard get_block_t calback function
2525  * when there is no desire to allocate any blocks.  It is used as a
2526  * callback function for block_prepare_write(), nobh_writepage(), and
2527  * block_write_full_page().  These functions should only try to map a
2528  * single block at a time.
2529  *
2530  * Since this function doesn't do block allocations even if the caller
2531  * requests it by passing in create=1, it is critically important that
2532  * any caller checks to make sure that any buffer heads are returned
2533  * by this function are either all already mapped or marked for
2534  * delayed allocation before calling nobh_writepage() or
2535  * block_write_full_page().  Otherwise, b_blocknr could be left
2536  * unitialized, and the page write functions will be taken by
2537  * surprise.
2538  */
2539 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
2540                                    struct buffer_head *bh_result, int create)
2541 {
2542         int ret = 0;
2543         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
2544
2545         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
2546
2547         /*
2548          * we don't want to do block allocation in writepage
2549          * so call get_block_wrap with create = 0
2550          */
2551         ret = ext4_get_blocks(NULL, inode, iblock, max_blocks, bh_result, 0);
2552         if (ret > 0) {
2553                 bh_result->b_size = (ret << inode->i_blkbits);
2554                 ret = 0;
2555         }
2556         return ret;
2557 }
2558
2559 static int bget_one(handle_t *handle, struct buffer_head *bh)
2560 {
2561         get_bh(bh);
2562         return 0;
2563 }
2564
2565 static int bput_one(handle_t *handle, struct buffer_head *bh)
2566 {
2567         put_bh(bh);
2568         return 0;
2569 }
2570
2571 static int __ext4_journalled_writepage(struct page *page,
2572                                        unsigned int len)
2573 {
2574         struct address_space *mapping = page->mapping;
2575         struct inode *inode = mapping->host;
2576         struct buffer_head *page_bufs;
2577         handle_t *handle = NULL;
2578         int ret = 0;
2579         int err;
2580
2581         page_bufs = page_buffers(page);
2582         BUG_ON(!page_bufs);
2583         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
2584         /* As soon as we unlock the page, it can go away, but we have
2585          * references to buffers so we are safe */
2586         unlock_page(page);
2587
2588         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
2589         if (IS_ERR(handle)) {
2590                 ret = PTR_ERR(handle);
2591                 goto out;
2592         }
2593
2594         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2595                                 do_journal_get_write_access);
2596
2597         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
2598                                 write_end_fn);
2599         if (ret == 0)
2600                 ret = err;
2601         err = ext4_journal_stop(handle);
2602         if (!ret)
2603                 ret = err;
2604
2605         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
2606         EXT4_I(inode)->i_state |= EXT4_STATE_JDATA;
2607 out:
2608         return ret;
2609 }
2610
2611 /*
2612  * Note that we don't need to start a transaction unless we're journaling data
2613  * because we should have holes filled from ext4_page_mkwrite(). We even don't
2614  * need to file the inode to the transaction's list in ordered mode because if
2615  * we are writing back data added by write(), the inode is already there and if
2616  * we are writing back data modified via mmap(), noone guarantees in which
2617  * transaction the data will hit the disk. In case we are journaling data, we
2618  * cannot start transaction directly because transaction start ranks above page
2619  * lock so we have to do some magic.
2620  *
2621  * This function can get called via...
2622  *   - ext4_da_writepages after taking page lock (have journal handle)
2623  *   - journal_submit_inode_data_buffers (no journal handle)
2624  *   - shrink_page_list via pdflush (no journal handle)
2625  *   - grab_page_cache when doing write_begin (have journal handle)
2626  *
2627  * We don't do any block allocation in this function. If we have page with
2628  * multiple blocks we need to write those buffer_heads that are mapped. This
2629  * is important for mmaped based write. So if we do with blocksize 1K
2630  * truncate(f, 1024);
2631  * a = mmap(f, 0, 4096);
2632  * a[0] = 'a';
2633  * truncate(f, 4096);
2634  * we have in the page first buffer_head mapped via page_mkwrite call back
2635  * but other bufer_heads would be unmapped but dirty(dirty done via the
2636  * do_wp_page). So writepage should write the first block. If we modify
2637  * the mmap area beyond 1024 we will again get a page_fault and the
2638  * page_mkwrite callback will do the block allocation and mark the
2639  * buffer_heads mapped.
2640  *
2641  * We redirty the page if we have any buffer_heads that is either delay or
2642  * unwritten in the page.
2643  *
2644  * We can get recursively called as show below.
2645  *
2646  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2647  *              ext4_writepage()
2648  *
2649  * But since we don't do any block allocation we should not deadlock.
2650  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2651  */
2652 static int ext4_writepage(struct page *page,
2653                           struct writeback_control *wbc)
2654 {
2655         int ret = 0;
2656         loff_t size;
2657         unsigned int len;
2658         struct buffer_head *page_bufs;
2659         struct inode *inode = page->mapping->host;
2660
2661         trace_ext4_writepage(inode, page);
2662         size = i_size_read(inode);
2663         if (page->index == size >> PAGE_CACHE_SHIFT)
2664                 len = size & ~PAGE_CACHE_MASK;
2665         else
2666                 len = PAGE_CACHE_SIZE;
2667
2668         if (page_has_buffers(page)) {
2669                 page_bufs = page_buffers(page);
2670                 if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2671                                         ext4_bh_delay_or_unwritten)) {
2672                         /*
2673                          * We don't want to do  block allocation
2674                          * So redirty the page and return
2675                          * We may reach here when we do a journal commit
2676                          * via journal_submit_inode_data_buffers.
2677                          * If we don't have mapping block we just ignore
2678                          * them. We can also reach here via shrink_page_list
2679                          */
2680                         redirty_page_for_writepage(wbc, page);
2681                         unlock_page(page);
2682                         return 0;
2683                 }
2684         } else {
2685                 /*
2686                  * The test for page_has_buffers() is subtle:
2687                  * We know the page is dirty but it lost buffers. That means
2688                  * that at some moment in time after write_begin()/write_end()
2689                  * has been called all buffers have been clean and thus they
2690                  * must have been written at least once. So they are all
2691                  * mapped and we can happily proceed with mapping them
2692                  * and writing the page.
2693                  *
2694                  * Try to initialize the buffer_heads and check whether
2695                  * all are mapped and non delay. We don't want to
2696                  * do block allocation here.
2697                  */
2698                 ret = block_prepare_write(page, 0, len,
2699                                           noalloc_get_block_write);
2700                 if (!ret) {
2701                         page_bufs = page_buffers(page);
2702                         /* check whether all are mapped and non delay */
2703                         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2704                                                 ext4_bh_delay_or_unwritten)) {
2705                                 redirty_page_for_writepage(wbc, page);
2706                                 unlock_page(page);
2707                                 return 0;
2708                         }
2709                 } else {
2710                         /*
2711                          * We can't do block allocation here
2712                          * so just redity the page and unlock
2713                          * and return
2714                          */
2715                         redirty_page_for_writepage(wbc, page);
2716                         unlock_page(page);
2717                         return 0;
2718                 }
2719                 /* now mark the buffer_heads as dirty and uptodate */
2720                 block_commit_write(page, 0, len);
2721         }
2722
2723         if (PageChecked(page) && ext4_should_journal_data(inode)) {
2724                 /*
2725                  * It's mmapped pagecache.  Add buffers and journal it.  There
2726                  * doesn't seem much point in redirtying the page here.
2727                  */
2728                 ClearPageChecked(page);
2729                 return __ext4_journalled_writepage(page, len);
2730         }
2731
2732         if (test_opt(inode->i_sb, NOBH) && ext4_should_writeback_data(inode))
2733                 ret = nobh_writepage(page, noalloc_get_block_write, wbc);
2734         else
2735                 ret = block_write_full_page(page, noalloc_get_block_write,
2736                                             wbc);
2737
2738         return ret;
2739 }
2740
2741 /*
2742  * This is called via ext4_da_writepages() to
2743  * calulate the total number of credits to reserve to fit
2744  * a single extent allocation into a single transaction,
2745  * ext4_da_writpeages() will loop calling this before
2746  * the block allocation.
2747  */
2748
2749 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2750 {
2751         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2752
2753         /*
2754          * With non-extent format the journal credit needed to
2755          * insert nrblocks contiguous block is dependent on
2756          * number of contiguous block. So we will limit
2757          * number of contiguous block to a sane value
2758          */
2759         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) &&
2760             (max_blocks > EXT4_MAX_TRANS_DATA))
2761                 max_blocks = EXT4_MAX_TRANS_DATA;
2762
2763         return ext4_chunk_trans_blocks(inode, max_blocks);
2764 }
2765
2766 static int ext4_da_writepages(struct address_space *mapping,
2767                               struct writeback_control *wbc)
2768 {
2769         pgoff_t index;
2770         int range_whole = 0;
2771         handle_t *handle = NULL;
2772         struct mpage_da_data mpd;
2773         struct inode *inode = mapping->host;
2774         int no_nrwrite_index_update;
2775         int pages_written = 0;
2776         long pages_skipped;
2777         unsigned int max_pages;
2778         int range_cyclic, cycled = 1, io_done = 0;
2779         int needed_blocks, ret = 0;
2780         long desired_nr_to_write, nr_to_writebump = 0;
2781         loff_t range_start = wbc->range_start;
2782         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2783
2784         trace_ext4_da_writepages(inode, wbc);
2785
2786         /*
2787          * No pages to write? This is mainly a kludge to avoid starting
2788          * a transaction for special inodes like journal inode on last iput()
2789          * because that could violate lock ordering on umount
2790          */
2791         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2792                 return 0;
2793
2794         /*
2795          * If the filesystem has aborted, it is read-only, so return
2796          * right away instead of dumping stack traces later on that
2797          * will obscure the real source of the problem.  We test
2798          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2799          * the latter could be true if the filesystem is mounted
2800          * read-only, and in that case, ext4_da_writepages should
2801          * *never* be called, so if that ever happens, we would want
2802          * the stack trace.
2803          */
2804         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2805                 return -EROFS;
2806
2807         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2808                 range_whole = 1;
2809
2810         range_cyclic = wbc->range_cyclic;
2811         if (wbc->range_cyclic) {
2812                 index = mapping->writeback_index;
2813                 if (index)
2814                         cycled = 0;
2815                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2816                 wbc->range_end  = LLONG_MAX;
2817                 wbc->range_cyclic = 0;
2818         } else
2819                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2820
2821         /*
2822          * This works around two forms of stupidity.  The first is in
2823          * the writeback code, which caps the maximum number of pages
2824          * written to be 1024 pages.  This is wrong on multiple
2825          * levels; different architectues have a different page size,
2826          * which changes the maximum amount of data which gets
2827          * written.  Secondly, 4 megabytes is way too small.  XFS
2828          * forces this value to be 16 megabytes by multiplying
2829          * nr_to_write parameter by four, and then relies on its
2830          * allocator to allocate larger extents to make them
2831          * contiguous.  Unfortunately this brings us to the second
2832          * stupidity, which is that ext4's mballoc code only allocates
2833          * at most 2048 blocks.  So we force contiguous writes up to
2834          * the number of dirty blocks in the inode, or
2835          * sbi->max_writeback_mb_bump whichever is smaller.
2836          */
2837         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2838         if (!range_cyclic && range_whole)
2839                 desired_nr_to_write = wbc->nr_to_write * 8;
2840         else
2841                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2842                                                            max_pages);
2843         if (desired_nr_to_write > max_pages)
2844                 desired_nr_to_write = max_pages;
2845
2846         if (wbc->nr_to_write < desired_nr_to_write) {
2847                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2848                 wbc->nr_to_write = desired_nr_to_write;
2849         }
2850
2851         mpd.wbc = wbc;
2852         mpd.inode = mapping->host;
2853
2854         /*
2855          * we don't want write_cache_pages to update
2856          * nr_to_write and writeback_index
2857          */
2858         no_nrwrite_index_update = wbc->no_nrwrite_index_update;
2859         wbc->no_nrwrite_index_update = 1;
2860         pages_skipped = wbc->pages_skipped;
2861
2862 retry:
2863         while (!ret && wbc->nr_to_write > 0) {
2864
2865                 /*
2866                  * we  insert one extent at a time. So we need
2867                  * credit needed for single extent allocation.
2868                  * journalled mode is currently not supported
2869                  * by delalloc
2870                  */
2871                 BUG_ON(ext4_should_journal_data(inode));
2872                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2873
2874                 /* start a new transaction*/
2875                 handle = ext4_journal_start(inode, needed_blocks);
2876                 if (IS_ERR(handle)) {
2877                         ret = PTR_ERR(handle);
2878                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2879                                "%ld pages, ino %lu; err %d\n", __func__,
2880                                 wbc->nr_to_write, inode->i_ino, ret);
2881                         goto out_writepages;
2882                 }
2883
2884                 /*
2885                  * Now call __mpage_da_writepage to find the next
2886                  * contiguous region of logical blocks that need
2887                  * blocks to be allocated by ext4.  We don't actually
2888                  * submit the blocks for I/O here, even though
2889                  * write_cache_pages thinks it will, and will set the
2890                  * pages as clean for write before calling
2891                  * __mpage_da_writepage().
2892                  */
2893                 mpd.b_size = 0;
2894                 mpd.b_state = 0;
2895                 mpd.b_blocknr = 0;
2896                 mpd.first_page = 0;
2897                 mpd.next_page = 0;
2898                 mpd.io_done = 0;
2899                 mpd.pages_written = 0;
2900                 mpd.retval = 0;
2901                 ret = write_cache_pages(mapping, wbc, __mpage_da_writepage,
2902                                         &mpd);
2903                 /*
2904                  * If we have a contiguous extent of pages and we
2905                  * haven't done the I/O yet, map the blocks and submit
2906                  * them for I/O.
2907                  */
2908                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2909                         if (mpage_da_map_blocks(&mpd) == 0)
2910                                 mpage_da_submit_io(&mpd);
2911                         mpd.io_done = 1;
2912                         ret = MPAGE_DA_EXTENT_TAIL;
2913                 }
2914                 trace_ext4_da_write_pages(inode, &mpd);
2915                 wbc->nr_to_write -= mpd.pages_written;
2916
2917                 ext4_journal_stop(handle);
2918
2919                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2920                         /* commit the transaction which would
2921                          * free blocks released in the transaction
2922                          * and try again
2923                          */
2924                         jbd2_journal_force_commit_nested(sbi->s_journal);
2925                         wbc->pages_skipped = pages_skipped;
2926                         ret = 0;
2927                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2928                         /*
2929                          * got one extent now try with
2930                          * rest of the pages
2931                          */
2932                         pages_written += mpd.pages_written;
2933                         wbc->pages_skipped = pages_skipped;
2934                         ret = 0;
2935                         io_done = 1;
2936                 } else if (wbc->nr_to_write)
2937                         /*
2938                          * There is no more writeout needed
2939                          * or we requested for a noblocking writeout
2940                          * and we found the device congested
2941                          */
2942                         break;
2943         }
2944         if (!io_done && !cycled) {
2945                 cycled = 1;
2946                 index = 0;
2947                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2948                 wbc->range_end  = mapping->writeback_index - 1;
2949                 goto retry;
2950         }
2951         if (pages_skipped != wbc->pages_skipped)
2952                 ext4_msg(inode->i_sb, KERN_CRIT,
2953                          "This should not happen leaving %s "
2954                          "with nr_to_write = %ld ret = %d\n",
2955                          __func__, wbc->nr_to_write, ret);
2956
2957         /* Update index */
2958         index += pages_written;
2959         wbc->range_cyclic = range_cyclic;
2960         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2961                 /*
2962                  * set the writeback_index so that range_cyclic
2963                  * mode will write it back later
2964                  */
2965                 mapping->writeback_index = index;
2966
2967 out_writepages:
2968         if (!no_nrwrite_index_update)
2969                 wbc->no_nrwrite_index_update = 0;
2970         if (wbc->nr_to_write > nr_to_writebump)
2971                 wbc->nr_to_write -= nr_to_writebump;
2972         wbc->range_start = range_start;
2973         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2974         return ret;
2975 }
2976
2977 #define FALL_BACK_TO_NONDELALLOC 1
2978 static int ext4_nonda_switch(struct super_block *sb)
2979 {
2980         s64 free_blocks, dirty_blocks;
2981         struct ext4_sb_info *sbi = EXT4_SB(sb);
2982
2983         /*
2984          * switch to non delalloc mode if we are running low
2985          * on free block. The free block accounting via percpu
2986          * counters can get slightly wrong with percpu_counter_batch getting
2987          * accumulated on each CPU without updating global counters
2988          * Delalloc need an accurate free block accounting. So switch
2989          * to non delalloc when we are near to error range.
2990          */
2991         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2992         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2993         if (2 * free_blocks < 3 * dirty_blocks ||
2994                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2995                 /*
2996                  * free block count is less than 150% of dirty blocks
2997                  * or free blocks is less than watermark
2998                  */
2999                 return 1;
3000         }
3001         /*
3002          * Even if we don't switch but are nearing capacity,
3003          * start pushing delalloc when 1/2 of free blocks are dirty.
3004          */
3005         if (free_blocks < 2 * dirty_blocks)
3006                 writeback_inodes_sb_if_idle(sb);
3007
3008         return 0;
3009 }
3010
3011 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
3012                                loff_t pos, unsigned len, unsigned flags,
3013                                struct page **pagep, void **fsdata)
3014 {
3015         int ret, retries = 0;
3016         struct page *page;
3017         pgoff_t index;
3018         unsigned from, to;
3019         struct inode *inode = mapping->host;
3020         handle_t *handle;
3021
3022         index = pos >> PAGE_CACHE_SHIFT;
3023         from = pos & (PAGE_CACHE_SIZE - 1);
3024         to = from + len;
3025
3026         if (ext4_nonda_switch(inode->i_sb)) {
3027                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
3028                 return ext4_write_begin(file, mapping, pos,
3029                                         len, flags, pagep, fsdata);
3030         }
3031         *fsdata = (void *)0;
3032         trace_ext4_da_write_begin(inode, pos, len, flags);
3033 retry:
3034         /*
3035          * With delayed allocation, we don't log the i_disksize update
3036          * if there is delayed block allocation. But we still need
3037          * to journalling the i_disksize update if writes to the end
3038          * of file which has an already mapped buffer.
3039          */
3040         handle = ext4_journal_start(inode, 1);
3041         if (IS_ERR(handle)) {
3042                 ret = PTR_ERR(handle);
3043                 goto out;
3044         }
3045         /* We cannot recurse into the filesystem as the transaction is already
3046          * started */
3047         flags |= AOP_FLAG_NOFS;
3048
3049         page = grab_cache_page_write_begin(mapping, index, flags);
3050         if (!page) {
3051                 ext4_journal_stop(handle);
3052                 ret = -ENOMEM;
3053                 goto out;
3054         }
3055         *pagep = page;
3056
3057         ret = block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
3058                                 ext4_da_get_block_prep);
3059         if (ret < 0) {
3060                 unlock_page(page);
3061                 ext4_journal_stop(handle);
3062                 page_cache_release(page);
3063                 /*
3064                  * block_write_begin may have instantiated a few blocks
3065                  * outside i_size.  Trim these off again. Don't need
3066                  * i_size_read because we hold i_mutex.
3067                  */
3068                 if (pos + len > inode->i_size)
3069                         ext4_truncate_failed_write(inode);
3070         }
3071
3072         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3073                 goto retry;
3074 out:
3075         return ret;
3076 }
3077
3078 /*
3079  * Check if we should update i_disksize
3080  * when write to the end of file but not require block allocation
3081  */
3082 static int ext4_da_should_update_i_disksize(struct page *page,
3083                                             unsigned long offset)
3084 {
3085         struct buffer_head *bh;
3086         struct inode *inode = page->mapping->host;
3087         unsigned int idx;
3088         int i;
3089
3090         bh = page_buffers(page);
3091         idx = offset >> inode->i_blkbits;
3092
3093         for (i = 0; i < idx; i++)
3094                 bh = bh->b_this_page;
3095
3096         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
3097                 return 0;
3098         return 1;
3099 }
3100
3101 static int ext4_da_write_end(struct file *file,
3102                              struct address_space *mapping,
3103                              loff_t pos, unsigned len, unsigned copied,
3104                              struct page *page, void *fsdata)
3105 {
3106         struct inode *inode = mapping->host;
3107         int ret = 0, ret2;
3108         handle_t *handle = ext4_journal_current_handle();
3109         loff_t new_i_size;
3110         unsigned long start, end;
3111         int write_mode = (int)(unsigned long)fsdata;
3112
3113         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
3114                 if (ext4_should_order_data(inode)) {
3115                         return ext4_ordered_write_end(file, mapping, pos,
3116                                         len, copied, page, fsdata);
3117                 } else if (ext4_should_writeback_data(inode)) {
3118                         return ext4_writeback_write_end(file, mapping, pos,
3119                                         len, copied, page, fsdata);
3120                 } else {
3121                         BUG();
3122                 }
3123         }
3124
3125         trace_ext4_da_write_end(inode, pos, len, copied);
3126         start = pos & (PAGE_CACHE_SIZE - 1);
3127         end = start + copied - 1;
3128
3129         /*
3130          * generic_write_end() will run mark_inode_dirty() if i_size
3131          * changes.  So let's piggyback the i_disksize mark_inode_dirty
3132          * into that.
3133          */
3134
3135         new_i_size = pos + copied;
3136         if (new_i_size > EXT4_I(inode)->i_disksize) {
3137                 if (ext4_da_should_update_i_disksize(page, end)) {
3138                         down_write(&EXT4_I(inode)->i_data_sem);
3139                         if (new_i_size > EXT4_I(inode)->i_disksize) {
3140                                 /*
3141                                  * Updating i_disksize when extending file
3142                                  * without needing block allocation
3143                                  */
3144                                 if (ext4_should_order_data(inode))
3145                                         ret = ext4_jbd2_file_inode(handle,
3146                                                                    inode);
3147
3148                                 EXT4_I(inode)->i_disksize = new_i_size;
3149                         }
3150                         up_write(&EXT4_I(inode)->i_data_sem);
3151                         /* We need to mark inode dirty even if
3152                          * new_i_size is less that inode->i_size
3153                          * bu greater than i_disksize.(hint delalloc)
3154                          */
3155                         ext4_mark_inode_dirty(handle, inode);
3156                 }
3157         }
3158         ret2 = generic_write_end(file, mapping, pos, len, copied,
3159                                                         page, fsdata);
3160         copied = ret2;
3161         if (ret2 < 0)
3162                 ret = ret2;
3163         ret2 = ext4_journal_stop(handle);
3164         if (!ret)
3165                 ret = ret2;
3166
3167         return ret ? ret : copied;
3168 }
3169
3170 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
3171 {
3172         /*
3173          * Drop reserved blocks
3174          */
3175         BUG_ON(!PageLocked(page));
3176         if (!page_has_buffers(page))
3177                 goto out;
3178
3179         ext4_da_page_release_reservation(page, offset);
3180
3181 out:
3182         ext4_invalidatepage(page, offset);
3183
3184         return;
3185 }
3186
3187 /*
3188  * Force all delayed allocation blocks to be allocated for a given inode.
3189  */
3190 int ext4_alloc_da_blocks(struct inode *inode)
3191 {
3192         trace_ext4_alloc_da_blocks(inode);
3193
3194         if (!EXT4_I(inode)->i_reserved_data_blocks &&
3195             !EXT4_I(inode)->i_reserved_meta_blocks)
3196                 return 0;
3197
3198         /*
3199          * We do something simple for now.  The filemap_flush() will
3200          * also start triggering a write of the data blocks, which is
3201          * not strictly speaking necessary (and for users of
3202          * laptop_mode, not even desirable).  However, to do otherwise
3203          * would require replicating code paths in:
3204          *
3205          * ext4_da_writepages() ->
3206          *    write_cache_pages() ---> (via passed in callback function)
3207          *        __mpage_da_writepage() -->
3208          *           mpage_add_bh_to_extent()
3209          *           mpage_da_map_blocks()
3210          *
3211          * The problem is that write_cache_pages(), located in
3212          * mm/page-writeback.c, marks pages clean in preparation for
3213          * doing I/O, which is not desirable if we're not planning on
3214          * doing I/O at all.
3215          *
3216          * We could call write_cache_pages(), and then redirty all of
3217          * the pages by calling redirty_page_for_writeback() but that
3218          * would be ugly in the extreme.  So instead we would need to
3219          * replicate parts of the code in the above functions,
3220          * simplifying them becuase we wouldn't actually intend to
3221          * write out the pages, but rather only collect contiguous
3222          * logical block extents, call the multi-block allocator, and
3223          * then update the buffer heads with the block allocations.
3224          *
3225          * For now, though, we'll cheat by calling filemap_flush(),
3226          * which will map the blocks, and start the I/O, but not
3227          * actually wait for the I/O to complete.
3228          */
3229         return filemap_flush(inode->i_mapping);
3230 }
3231
3232 /*
3233  * bmap() is special.  It gets used by applications such as lilo and by
3234  * the swapper to find the on-disk block of a specific piece of data.
3235  *
3236  * Naturally, this is dangerous if the block concerned is still in the
3237  * journal.  If somebody makes a swapfile on an ext4 data-journaling
3238  * filesystem and enables swap, then they may get a nasty shock when the
3239  * data getting swapped to that swapfile suddenly gets overwritten by
3240  * the original zero's written out previously to the journal and
3241  * awaiting writeback in the kernel's buffer cache.
3242  *
3243  * So, if we see any bmap calls here on a modified, data-journaled file,
3244  * take extra steps to flush any blocks which might be in the cache.
3245  */
3246 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
3247 {
3248         struct inode *inode = mapping->host;
3249         journal_t *journal;
3250         int err;
3251
3252         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
3253                         test_opt(inode->i_sb, DELALLOC)) {
3254                 /*
3255                  * With delalloc we want to sync the file
3256                  * so that we can make sure we allocate
3257                  * blocks for file
3258                  */
3259                 filemap_write_and_wait(mapping);
3260         }
3261
3262         if (EXT4_JOURNAL(inode) && EXT4_I(inode)->i_state & EXT4_STATE_JDATA) {
3263                 /*
3264                  * This is a REALLY heavyweight approach, but the use of
3265                  * bmap on dirty files is expected to be extremely rare:
3266                  * only if we run lilo or swapon on a freshly made file
3267                  * do we expect this to happen.
3268                  *
3269                  * (bmap requires CAP_SYS_RAWIO so this does not
3270                  * represent an unprivileged user DOS attack --- we'd be
3271                  * in trouble if mortal users could trigger this path at
3272                  * will.)
3273                  *
3274                  * NB. EXT4_STATE_JDATA is not set on files other than
3275                  * regular files.  If somebody wants to bmap a directory
3276                  * or symlink and gets confused because the buffer
3277                  * hasn't yet been flushed to disk, they deserve
3278                  * everything they get.
3279                  */
3280
3281                 EXT4_I(inode)->i_state &= ~EXT4_STATE_JDATA;
3282                 journal = EXT4_JOURNAL(inode);
3283                 jbd2_journal_lock_updates(journal);
3284                 err = jbd2_journal_flush(journal);
3285                 jbd2_journal_unlock_updates(journal);
3286
3287                 if (err)
3288                         return 0;
3289         }
3290
3291         return generic_block_bmap(mapping, block, ext4_get_block);
3292 }
3293
3294 static int ext4_readpage(struct file *file, struct page *page)
3295 {
3296         return mpage_readpage(page, ext4_get_block);
3297 }
3298
3299 static int
3300 ext4_readpages(struct file *file, struct address_space *mapping,
3301                 struct list_head *pages, unsigned nr_pages)
3302 {
3303         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
3304 }
3305
3306 static void ext4_invalidatepage(struct page *page, unsigned long offset)
3307 {
3308         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3309
3310         /*
3311          * If it's a full truncate we just forget about the pending dirtying
3312          */
3313         if (offset == 0)
3314                 ClearPageChecked(page);
3315
3316         if (journal)
3317                 jbd2_journal_invalidatepage(journal, page, offset);
3318         else
3319                 block_invalidatepage(page, offset);
3320 }
3321
3322 static int ext4_releasepage(struct page *page, gfp_t wait)
3323 {
3324         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
3325
3326         WARN_ON(PageChecked(page));
3327         if (!page_has_buffers(page))
3328                 return 0;
3329         if (journal)
3330                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
3331         else
3332                 return try_to_free_buffers(page);
3333 }
3334
3335 /*
3336  * O_DIRECT for ext3 (or indirect map) based files
3337  *
3338  * If the O_DIRECT write will extend the file then add this inode to the
3339  * orphan list.  So recovery will truncate it back to the original size
3340  * if the machine crashes during the write.
3341  *
3342  * If the O_DIRECT write is intantiating holes inside i_size and the machine
3343  * crashes then stale disk data _may_ be exposed inside the file. But current
3344  * VFS code falls back into buffered path in that case so we are safe.
3345  */
3346 static ssize_t ext4_ind_direct_IO(int rw, struct kiocb *iocb,
3347                               const struct iovec *iov, loff_t offset,
3348                               unsigned long nr_segs)
3349 {
3350         struct file *file = iocb->ki_filp;
3351         struct inode *inode = file->f_mapping->host;
3352         struct ext4_inode_info *ei = EXT4_I(inode);
3353         handle_t *handle;
3354         ssize_t ret;
3355         int orphan = 0;
3356         size_t count = iov_length(iov, nr_segs);
3357         int retries = 0;
3358
3359         if (rw == WRITE) {
3360                 loff_t final_size = offset + count;
3361
3362                 if (final_size > inode->i_size) {
3363                         /* Credits for sb + inode write */
3364                         handle = ext4_journal_start(inode, 2);
3365                         if (IS_ERR(handle)) {
3366                                 ret = PTR_ERR(handle);
3367                                 goto out;
3368                         }
3369                         ret = ext4_orphan_add(handle, inode);
3370                         if (ret) {
3371                                 ext4_journal_stop(handle);
3372                                 goto out;
3373                         }
3374                         orphan = 1;
3375                         ei->i_disksize = inode->i_size;
3376                         ext4_journal_stop(handle);
3377                 }
3378         }
3379
3380 retry:
3381         ret = blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev, iov,
3382                                  offset, nr_segs,
3383                                  ext4_get_block, NULL);
3384         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
3385                 goto retry;
3386
3387         if (orphan) {
3388                 int err;
3389
3390                 /* Credits for sb + inode write */
3391                 handle = ext4_journal_start(inode, 2);
3392                 if (IS_ERR(handle)) {
3393                         /* This is really bad luck. We've written the data
3394                          * but cannot extend i_size. Bail out and pretend
3395                          * the write failed... */
3396                         ret = PTR_ERR(handle);
3397                         goto out;
3398                 }
3399                 if (inode->i_nlink)
3400                         ext4_orphan_del(handle, inode);
3401                 if (ret > 0) {
3402                         loff_t end = offset + ret;
3403                         if (end > inode->i_size) {
3404                                 ei->i_disksize = end;
3405                                 i_size_write(inode, end);
3406                                 /*
3407                                  * We're going to return a positive `ret'
3408                                  * here due to non-zero-length I/O, so there's
3409                                  * no way of reporting error returns from
3410                                  * ext4_mark_inode_dirty() to userspace.  So
3411                                  * ignore it.
3412                                  */
3413                                 ext4_mark_inode_dirty(handle, inode);
3414                         }
3415                 }
3416                 err = ext4_journal_stop(handle);
3417                 if (ret == 0)
3418                         ret = err;
3419         }
3420 out:
3421         return ret;
3422 }
3423
3424 static int ext4_get_block_dio_write(struct inode *inode, sector_t iblock,
3425                    struct buffer_head *bh_result, int create)
3426 {
3427         handle_t *handle = NULL;
3428         int ret = 0;
3429         unsigned max_blocks = bh_result->b_size >> inode->i_blkbits;
3430         int dio_credits;
3431
3432         ext4_debug("ext4_get_block_dio_write: inode %lu, create flag %d\n",
3433                    inode->i_ino, create);
3434         /*
3435          * DIO VFS code passes create = 0 flag for write to
3436          * the middle of file. It does this to avoid block
3437          * allocation for holes, to prevent expose stale data
3438          * out when there is parallel buffered read (which does
3439          * not hold the i_mutex lock) while direct IO write has
3440          * not completed. DIO request on holes finally falls back
3441          * to buffered IO for this reason.
3442          *
3443          * For ext4 extent based file, since we support fallocate,
3444          * new allocated extent as uninitialized, for holes, we
3445          * could fallocate blocks for holes, thus parallel
3446          * buffered IO read will zero out the page when read on
3447          * a hole while parallel DIO write to the hole has not completed.
3448          *
3449          * when we come here, we know it's a direct IO write to
3450          * to the middle of file (<i_size)
3451          * so it's safe to override the create flag from VFS.
3452          */
3453         create = EXT4_GET_BLOCKS_DIO_CREATE_EXT;
3454
3455         if (max_blocks > DIO_MAX_BLOCKS)
3456                 max_blocks = DIO_MAX_BLOCKS;
3457         dio_credits = ext4_chunk_trans_blocks(inode, max_blocks);
3458         handle = ext4_journal_start(inode, dio_credits);
3459         if (IS_ERR(handle)) {
3460                 ret = PTR_ERR(handle);
3461                 goto out;
3462         }
3463         ret = ext4_get_blocks(handle, inode, iblock, max_blocks, bh_result,
3464                               create);
3465         if (ret > 0) {
3466                 bh_result->b_size = (ret << inode->i_blkbits);
3467                 ret = 0;
3468         }
3469         ext4_journal_stop(handle);
3470 out:
3471         return ret;
3472 }
3473
3474 static void ext4_free_io_end(ext4_io_end_t *io)
3475 {
3476         BUG_ON(!io);
3477         iput(io->inode);
3478         kfree(io);
3479 }
3480 static void dump_aio_dio_list(struct inode * inode)
3481 {
3482 #ifdef  EXT4_DEBUG
3483         struct list_head *cur, *before, *after;
3484         ext4_io_end_t *io, *io0, *io1;
3485
3486         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3487                 ext4_debug("inode %lu aio dio list is empty\n", inode->i_ino);
3488                 return;
3489         }
3490
3491         ext4_debug("Dump inode %lu aio_dio_completed_IO list \n", inode->i_ino);
3492         list_for_each_entry(io, &EXT4_I(inode)->i_aio_dio_complete_list, list){
3493                 cur = &io->list;
3494                 before = cur->prev;
3495                 io0 = container_of(before, ext4_io_end_t, list);
3496                 after = cur->next;
3497                 io1 = container_of(after, ext4_io_end_t, list);
3498
3499                 ext4_debug("io 0x%p from inode %lu,prev 0x%p,next 0x%p\n",
3500                             io, inode->i_ino, io0, io1);
3501         }
3502 #endif
3503 }
3504
3505 /*
3506  * check a range of space and convert unwritten extents to written.
3507  */
3508 static int ext4_end_aio_dio_nolock(ext4_io_end_t *io)
3509 {
3510         struct inode *inode = io->inode;
3511         loff_t offset = io->offset;
3512         size_t size = io->size;
3513         int ret = 0;
3514
3515         ext4_debug("end_aio_dio_onlock: io 0x%p from inode %lu,list->next 0x%p,"
3516                    "list->prev 0x%p\n",
3517                    io, inode->i_ino, io->list.next, io->list.prev);
3518
3519         if (list_empty(&io->list))
3520                 return ret;
3521
3522         if (io->flag != DIO_AIO_UNWRITTEN)
3523                 return ret;
3524
3525         if (offset + size <= i_size_read(inode))
3526                 ret = ext4_convert_unwritten_extents(inode, offset, size);
3527
3528         if (ret < 0) {
3529                 printk(KERN_EMERG "%s: failed to convert unwritten"
3530                         "extents to written extents, error is %d"
3531                         " io is still on inode %lu aio dio list\n",
3532                        __func__, ret, inode->i_ino);
3533                 return ret;
3534         }
3535
3536         /* clear the DIO AIO unwritten flag */
3537         io->flag = 0;
3538         return ret;
3539 }
3540 /*
3541  * work on completed aio dio IO, to convert unwritten extents to extents
3542  */
3543 static void ext4_end_aio_dio_work(struct work_struct *work)
3544 {
3545         ext4_io_end_t *io  = container_of(work, ext4_io_end_t, work);
3546         struct inode *inode = io->inode;
3547         int ret = 0;
3548
3549         mutex_lock(&inode->i_mutex);
3550         ret = ext4_end_aio_dio_nolock(io);
3551         if (ret >= 0) {
3552                 if (!list_empty(&io->list))
3553                         list_del_init(&io->list);
3554                 ext4_free_io_end(io);
3555         }
3556         mutex_unlock(&inode->i_mutex);
3557 }
3558 /*
3559  * This function is called from ext4_sync_file().
3560  *
3561  * When AIO DIO IO is completed, the work to convert unwritten
3562  * extents to written is queued on workqueue but may not get immediately
3563  * scheduled. When fsync is called, we need to ensure the
3564  * conversion is complete before fsync returns.
3565  * The inode keeps track of a list of completed AIO from DIO path
3566  * that might needs to do the conversion. This function walks through
3567  * the list and convert the related unwritten extents to written.
3568  */
3569 int flush_aio_dio_completed_IO(struct inode *inode)
3570 {
3571         ext4_io_end_t *io;
3572         int ret = 0;
3573         int ret2 = 0;
3574
3575         if (list_empty(&EXT4_I(inode)->i_aio_dio_complete_list))
3576                 return ret;
3577
3578         dump_aio_dio_list(inode);
3579         while (!list_empty(&EXT4_I(inode)->i_aio_dio_complete_list)){
3580                 io = list_entry(EXT4_I(inode)->i_aio_dio_complete_list.next,
3581                                 ext4_io_end_t, list);
3582                 /*
3583                  * Calling ext4_end_aio_dio_nolock() to convert completed
3584                  * IO to written.
3585                  *
3586                  * When ext4_sync_file() is called, run_queue() may already
3587                  * about to flush the work corresponding to this io structure.
3588                  * It will be upset if it founds the io structure related
3589                  * to the work-to-be schedule is freed.
3590                  *
3591                  * Thus we need to keep the io structure still valid here after
3592                  * convertion finished. The io structure has a flag to
3593                  * avoid double converting from both fsync and background work
3594                  * queue work.
3595                  */
3596                 ret = ext4_end_aio_dio_nolock(io);
3597                 if (ret < 0)
3598                         ret2 = ret;
3599                 else
3600                         list_del_init(&io->list);
3601         }
3602         return (ret2 < 0) ? ret2 : 0;
3603 }
3604
3605 static ext4_io_end_t *ext4_init_io_end (struct inode *inode)
3606 {
3607         ext4_io_end_t *io = NULL;
3608
3609         io = kmalloc(sizeof(*io), GFP_NOFS);
3610
3611         if (io) {
3612                 igrab(inode);
3613                 io->inode = inode;
3614                 io->flag = 0;
3615                 io->offset = 0;
3616                 io->size = 0;
3617                 io->error = 0;
3618                 INIT_WORK(&io->work, ext4_end_aio_dio_work);
3619                 INIT_LIST_HEAD(&io->list);
3620         }
3621
3622         return io;
3623 }
3624
3625 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
3626                             ssize_t size, void *private)
3627 {
3628         ext4_io_end_t *io_end = iocb->private;
3629         struct workqueue_struct *wq;
3630
3631         /* if not async direct IO or dio with 0 bytes write, just return */
3632         if (!io_end || !size)
3633                 return;
3634
3635         ext_debug("ext4_end_io_dio(): io_end 0x%p"
3636                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
3637                   iocb->private, io_end->inode->i_ino, iocb, offset,
3638                   size);
3639
3640         /* if not aio dio with unwritten extents, just free io and return */
3641         if (io_end->flag != DIO_AIO_UNWRITTEN){
3642                 ext4_free_io_end(io_end);
3643                 iocb->private = NULL;
3644                 return;
3645         }
3646
3647         io_end->offset = offset;
3648         io_end->size = size;
3649         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
3650
3651         /* queue the work to convert unwritten extents to written */
3652         queue_work(wq, &io_end->work);
3653
3654         /* Add the io_end to per-inode completed aio dio list*/
3655         list_add_tail(&io_end->list,
3656                  &EXT4_I(io_end->inode)->i_aio_dio_complete_list);
3657         iocb->private = NULL;
3658 }
3659 /*
3660  * For ext4 extent files, ext4 will do direct-io write to holes,
3661  * preallocated extents, and those write extend the file, no need to
3662  * fall back to buffered IO.
3663  *
3664  * For holes, we fallocate those blocks, mark them as unintialized
3665  * If those blocks were preallocated, we mark sure they are splited, but
3666  * still keep the range to write as unintialized.
3667  *
3668  * The unwrritten extents will be converted to written when DIO is completed.
3669  * For async direct IO, since the IO may still pending when return, we
3670  * set up an end_io call back function, which will do the convertion
3671  * when async direct IO completed.
3672  *
3673  * If the O_DIRECT write will extend the file then add this inode to the
3674  * orphan list.  So recovery will truncate it back to the original size
3675  * if the machine crashes during the write.
3676  *
3677  */
3678 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3679                               const struct iovec *iov, loff_t offset,
3680                               unsigned long nr_segs)
3681 {
3682         struct file *file = iocb->ki_filp;
3683         struct inode *inode = file->f_mapping->host;
3684         ssize_t ret;
3685         size_t count = iov_length(iov, nr_segs);
3686
3687         loff_t final_size = offset + count;
3688         if (rw == WRITE && final_size <= inode->i_size) {
3689                 /*
3690                  * We could direct write to holes and fallocate.
3691                  *
3692                  * Allocated blocks to fill the hole are marked as uninitialized
3693                  * to prevent paralel buffered read to expose the stale data
3694                  * before DIO complete the data IO.
3695                  *
3696                  * As to previously fallocated extents, ext4 get_block
3697                  * will just simply mark the buffer mapped but still
3698                  * keep the extents uninitialized.
3699                  *
3700                  * for non AIO case, we will convert those unwritten extents
3701                  * to written after return back from blockdev_direct_IO.
3702                  *
3703                  * for async DIO, the conversion needs to be defered when
3704                  * the IO is completed. The ext4 end_io callback function
3705                  * will be called to take care of the conversion work.
3706                  * Here for async case, we allocate an io_end structure to
3707                  * hook to the iocb.
3708                  */
3709                 iocb->private = NULL;
3710                 EXT4_I(inode)->cur_aio_dio = NULL;
3711                 if (!is_sync_kiocb(iocb)) {
3712                         iocb->private = ext4_init_io_end(inode);
3713                         if (!iocb->private)
3714                                 return -ENOMEM;
3715                         /*
3716                          * we save the io structure for current async
3717                          * direct IO, so that later ext4_get_blocks()
3718                          * could flag the io structure whether there
3719                          * is a unwritten extents needs to be converted
3720                          * when IO is completed.
3721                          */
3722                         EXT4_I(inode)->cur_aio_dio = iocb->private;
3723                 }
3724
3725                 ret = blockdev_direct_IO(rw, iocb, inode,
3726                                          inode->i_sb->s_bdev, iov,
3727                                          offset, nr_segs,
3728                                          ext4_get_block_dio_write,
3729                                          ext4_end_io_dio);
3730                 if (iocb->private)
3731                         EXT4_I(inode)->cur_aio_dio = NULL;
3732                 /*
3733                  * The io_end structure takes a reference to the inode,
3734                  * that structure needs to be destroyed and the
3735                  * reference to the inode need to be dropped, when IO is
3736                  * complete, even with 0 byte write, or failed.
3737                  *
3738                  * In the successful AIO DIO case, the io_end structure will be
3739                  * desctroyed and the reference to the inode will be dropped
3740                  * after the end_io call back function is called.
3741                  *
3742                  * In the case there is 0 byte write, or error case, since
3743                  * VFS direct IO won't invoke the end_io call back function,
3744                  * we need to free the end_io structure here.
3745                  */
3746                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
3747                         ext4_free_io_end(iocb->private);
3748                         iocb->private = NULL;
3749                 } else if (ret > 0 && (EXT4_I(inode)->i_state &
3750                                        EXT4_STATE_DIO_UNWRITTEN)) {
3751                         int err;
3752                         /*
3753                          * for non AIO case, since the IO is already
3754                          * completed, we could do the convertion right here
3755                          */
3756                         err = ext4_convert_unwritten_extents(inode,
3757                                                              offset, ret);
3758                         if (err < 0)
3759                                 ret = err;
3760                         EXT4_I(inode)->i_state &= ~EXT4_STATE_DIO_UNWRITTEN;
3761                 }
3762                 return ret;
3763         }
3764
3765         /* for write the the end of file case, we fall back to old way */
3766         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3767 }
3768
3769 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
3770                               const struct iovec *iov, loff_t offset,
3771                               unsigned long nr_segs)
3772 {
3773         struct file *file = iocb->ki_filp;
3774         struct inode *inode = file->f_mapping->host;
3775
3776         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)
3777                 return ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3778
3779         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3780 }
3781
3782 /*
3783  * Pages can be marked dirty completely asynchronously from ext4's journalling
3784  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3785  * much here because ->set_page_dirty is called under VFS locks.  The page is
3786  * not necessarily locked.
3787  *
3788  * We cannot just dirty the page and leave attached buffers clean, because the
3789  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3790  * or jbddirty because all the journalling code will explode.
3791  *
3792  * So what we do is to mark the page "pending dirty" and next time writepage
3793  * is called, propagate that into the buffers appropriately.
3794  */
3795 static int ext4_journalled_set_page_dirty(struct page *page)
3796 {
3797         SetPageChecked(page);
3798         return __set_page_dirty_nobuffers(page);
3799 }
3800
3801 static const struct address_space_operations ext4_ordered_aops = {
3802         .readpage               = ext4_readpage,
3803         .readpages              = ext4_readpages,
3804         .writepage              = ext4_writepage,
3805         .sync_page              = block_sync_page,
3806         .write_begin            = ext4_write_begin,
3807         .write_end              = ext4_ordered_write_end,
3808         .bmap                   = ext4_bmap,
3809         .invalidatepage         = ext4_invalidatepage,
3810         .releasepage            = ext4_releasepage,
3811         .direct_IO              = ext4_direct_IO,
3812         .migratepage            = buffer_migrate_page,
3813         .is_partially_uptodate  = block_is_partially_uptodate,
3814         .error_remove_page      = generic_error_remove_page,
3815 };
3816
3817 static const struct address_space_operations ext4_writeback_aops = {
3818         .readpage               = ext4_readpage,
3819         .readpages              = ext4_readpages,
3820         .writepage              = ext4_writepage,
3821         .sync_page              = block_sync_page,
3822         .write_begin            = ext4_write_begin,
3823         .write_end              = ext4_writeback_write_end,
3824         .bmap                   = ext4_bmap,
3825         .invalidatepage         = ext4_invalidatepage,
3826         .releasepage            = ext4_releasepage,
3827         .direct_IO              = ext4_direct_IO,
3828         .migratepage            = buffer_migrate_page,
3829         .is_partially_uptodate  = block_is_partially_uptodate,
3830         .error_remove_page      = generic_error_remove_page,
3831 };
3832
3833 static const struct address_space_operations ext4_journalled_aops = {
3834         .readpage               = ext4_readpage,
3835         .readpages              = ext4_readpages,
3836         .writepage              = ext4_writepage,
3837         .sync_page              = block_sync_page,
3838         .write_begin            = ext4_write_begin,
3839         .write_end              = ext4_journalled_write_end,
3840         .set_page_dirty         = ext4_journalled_set_page_dirty,
3841         .bmap                   = ext4_bmap,
3842         .invalidatepage         = ext4_invalidatepage,
3843         .releasepage            = ext4_releasepage,
3844         .is_partially_uptodate  = block_is_partially_uptodate,
3845         .error_remove_page      = generic_error_remove_page,
3846 };
3847
3848 static const struct address_space_operations ext4_da_aops = {
3849         .readpage               = ext4_readpage,
3850         .readpages              = ext4_readpages,
3851         .writepage              = ext4_writepage,
3852         .writepages             = ext4_da_writepages,
3853         .sync_page              = block_sync_page,
3854         .write_begin            = ext4_da_write_begin,
3855         .write_end              = ext4_da_write_end,
3856         .bmap                   = ext4_bmap,
3857         .invalidatepage         = ext4_da_invalidatepage,
3858         .releasepage            = ext4_releasepage,
3859         .direct_IO              = ext4_direct_IO,
3860         .migratepage            = buffer_migrate_page,
3861         .is_partially_uptodate  = block_is_partially_uptodate,
3862         .error_remove_page      = generic_error_remove_page,
3863 };
3864
3865 void ext4_set_aops(struct inode *inode)
3866 {
3867         if (ext4_should_order_data(inode) &&
3868                 test_opt(inode->i_sb, DELALLOC))
3869                 inode->i_mapping->a_ops = &ext4_da_aops;
3870         else if (ext4_should_order_data(inode))
3871                 inode->i_mapping->a_ops = &ext4_ordered_aops;
3872         else if (ext4_should_writeback_data(inode) &&
3873                  test_opt(inode->i_sb, DELALLOC))
3874                 inode->i_mapping->a_ops = &ext4_da_aops;
3875         else if (ext4_should_writeback_data(inode))
3876                 inode->i_mapping->a_ops = &ext4_writeback_aops;
3877         else
3878                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3879 }
3880
3881 /*
3882  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
3883  * up to the end of the block which corresponds to `from'.
3884  * This required during truncate. We need to physically zero the tail end
3885  * of that block so it doesn't yield old data if the file is later grown.
3886  */
3887 int ext4_block_truncate_page(handle_t *handle,
3888                 struct address_space *mapping, loff_t from)
3889 {
3890         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3891         unsigned offset = from & (PAGE_CACHE_SIZE-1);
3892         unsigned blocksize, length, pos;
3893         ext4_lblk_t iblock;
3894         struct inode *inode = mapping->host;
3895         struct buffer_head *bh;
3896         struct page *page;
3897         int err = 0;
3898
3899         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3900                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3901         if (!page)
3902                 return -EINVAL;
3903
3904         blocksize = inode->i_sb->s_blocksize;
3905         length = blocksize - (offset & (blocksize - 1));
3906         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3907
3908         /*
3909          * For "nobh" option,  we can only work if we don't need to
3910          * read-in the page - otherwise we create buffers to do the IO.
3911          */
3912         if (!page_has_buffers(page) && test_opt(inode->i_sb, NOBH) &&
3913              ext4_should_writeback_data(inode) && PageUptodate(page)) {
3914                 zero_user(page, offset, length);
3915                 set_page_dirty(page);
3916                 goto unlock;
3917         }
3918
3919         if (!page_has_buffers(page))
3920                 create_empty_buffers(page, blocksize, 0);
3921
3922         /* Find the buffer that contains "offset" */
3923         bh = page_buffers(page);
3924         pos = blocksize;
3925         while (offset >= pos) {
3926                 bh = bh->b_this_page;
3927                 iblock++;
3928                 pos += blocksize;
3929         }
3930
3931         err = 0;
3932         if (buffer_freed(bh)) {
3933                 BUFFER_TRACE(bh, "freed: skip");
3934                 goto unlock;
3935         }
3936
3937         if (!buffer_mapped(bh)) {
3938                 BUFFER_TRACE(bh, "unmapped");
3939                 ext4_get_block(inode, iblock, bh, 0);
3940                 /* unmapped? It's a hole - nothing to do */
3941                 if (!buffer_mapped(bh)) {
3942                         BUFFER_TRACE(bh, "still unmapped");
3943                         goto unlock;
3944                 }
3945         }
3946
3947         /* Ok, it's mapped. Make sure it's up-to-date */
3948         if (PageUptodate(page))
3949                 set_buffer_uptodate(bh);
3950
3951         if (!buffer_uptodate(bh)) {
3952                 err = -EIO;
3953                 ll_rw_block(READ, 1, &bh);
3954                 wait_on_buffer(bh);
3955                 /* Uhhuh. Read error. Complain and punt. */
3956                 if (!buffer_uptodate(bh))
3957                         goto unlock;
3958         }
3959
3960         if (ext4_should_journal_data(inode)) {
3961                 BUFFER_TRACE(bh, "get write access");
3962                 err = ext4_journal_get_write_access(handle, bh);
3963                 if (err)
3964                         goto unlock;
3965         }
3966
3967         zero_user(page, offset, length);
3968
3969         BUFFER_TRACE(bh, "zeroed end of block");
3970
3971         err = 0;
3972         if (ext4_should_journal_data(inode)) {
3973                 err = ext4_handle_dirty_metadata(handle, inode, bh);
3974         } else {
3975                 if (ext4_should_order_data(inode))
3976                         err = ext4_jbd2_file_inode(handle, inode);
3977                 mark_buffer_dirty(bh);
3978         }
3979
3980 unlock:
3981         unlock_page(page);
3982         page_cache_release(page);
3983         return err;
3984 }
3985
3986 /*
3987  * Probably it should be a library function... search for first non-zero word
3988  * or memcmp with zero_page, whatever is better for particular architecture.
3989  * Linus?
3990  */
3991 static inline int all_zeroes(__le32 *p, __le32 *q)
3992 {
3993         while (p < q)
3994                 if (*p++)
3995                         return 0;
3996         return 1;
3997 }
3998
3999 /**
4000  *      ext4_find_shared - find the indirect blocks for partial truncation.
4001  *      @inode:   inode in question
4002  *      @depth:   depth of the affected branch
4003  *      @offsets: offsets of pointers in that branch (see ext4_block_to_path)
4004  *      @chain:   place to store the pointers to partial indirect blocks
4005  *      @top:     place to the (detached) top of branch
4006  *
4007  *      This is a helper function used by ext4_truncate().
4008  *
4009  *      When we do truncate() we may have to clean the ends of several
4010  *      indirect blocks but leave the blocks themselves alive. Block is
4011  *      partially truncated if some data below the new i_size is refered
4012  *      from it (and it is on the path to the first completely truncated
4013  *      data block, indeed).  We have to free the top of that path along
4014  *      with everything to the right of the path. Since no allocation
4015  *      past the truncation point is possible until ext4_truncate()
4016  *      finishes, we may safely do the latter, but top of branch may
4017  *      require special attention - pageout below the truncation point
4018  *      might try to populate it.
4019  *
4020  *      We atomically detach the top of branch from the tree, store the
4021  *      block number of its root in *@top, pointers to buffer_heads of
4022  *      partially truncated blocks - in @chain[].bh and pointers to
4023  *      their last elements that should not be removed - in
4024  *      @chain[].p. Return value is the pointer to last filled element
4025  *      of @chain.
4026  *
4027  *      The work left to caller to do the actual freeing of subtrees:
4028  *              a) free the subtree starting from *@top
4029  *              b) free the subtrees whose roots are stored in
4030  *                      (@chain[i].p+1 .. end of @chain[i].bh->b_data)
4031  *              c) free the subtrees growing from the inode past the @chain[0].
4032  *                      (no partially truncated stuff there).  */
4033
4034 static Indirect *ext4_find_shared(struct inode *inode, int depth,
4035                                   ext4_lblk_t offsets[4], Indirect chain[4],
4036                                   __le32 *top)
4037 {
4038         Indirect *partial, *p;
4039         int k, err;
4040
4041         *top = 0;
4042         /* Make k index the deepest non-null offset + 1 */
4043         for (k = depth; k > 1 && !offsets[k-1]; k--)
4044                 ;
4045         partial = ext4_get_branch(inode, k, offsets, chain, &err);
4046         /* Writer: pointers */
4047         if (!partial)
4048                 partial = chain + k-1;
4049         /*
4050          * If the branch acquired continuation since we've looked at it -
4051          * fine, it should all survive and (new) top doesn't belong to us.
4052          */
4053         if (!partial->key && *partial->p)
4054                 /* Writer: end */
4055                 goto no_top;
4056         for (p = partial; (p > chain) && all_zeroes((__le32 *) p->bh->b_data, p->p); p--)
4057                 ;
4058         /*
4059          * OK, we've found the last block that must survive. The rest of our
4060          * branch should be detached before unlocking. However, if that rest
4061          * of branch is all ours and does not grow immediately from the inode
4062          * it's easier to cheat and just decrement partial->p.
4063          */
4064         if (p == chain + k - 1 && p > chain) {
4065                 p->p--;
4066         } else {
4067                 *top = *p->p;
4068                 /* Nope, don't do this in ext4.  Must leave the tree intact */
4069 #if 0
4070                 *p->p = 0;
4071 #endif
4072         }
4073         /* Writer: end */
4074
4075         while (partial > p) {
4076                 brelse(partial->bh);
4077                 partial--;
4078         }
4079 no_top:
4080         return partial;
4081 }
4082
4083 /*
4084  * Zero a number of block pointers in either an inode or an indirect block.
4085  * If we restart the transaction we must again get write access to the
4086  * indirect block for further modification.
4087  *
4088  * We release `count' blocks on disk, but (last - first) may be greater
4089  * than `count' because there can be holes in there.
4090  */
4091 static void ext4_clear_blocks(handle_t *handle, struct inode *inode,
4092                               struct buffer_head *bh,
4093                               ext4_fsblk_t block_to_free,
4094                               unsigned long count, __le32 *first,
4095                               __le32 *last)
4096 {
4097         __le32 *p;
4098         int     flags = EXT4_FREE_BLOCKS_FORGET;
4099
4100         if (S_ISDIR(inode->i_mode) || S_ISLNK(inode->i_mode))
4101                 flags |= EXT4_FREE_BLOCKS_METADATA;
4102
4103         if (try_to_extend_transaction(handle, inode)) {
4104                 if (bh) {
4105                         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
4106                         ext4_handle_dirty_metadata(handle, inode, bh);
4107                 }
4108                 ext4_mark_inode_dirty(handle, inode);
4109                 ext4_truncate_restart_trans(handle, inode,
4110                                             blocks_for_truncate(inode));
4111                 if (bh) {
4112                         BUFFER_TRACE(bh, "retaking write access");
4113                         ext4_journal_get_write_access(handle, bh);
4114                 }
4115         }
4116
4117         for (p = first; p < last; p++)
4118                 *p = 0;
4119
4120         ext4_free_blocks(handle, inode, 0, block_to_free, count, flags);
4121 }
4122
4123 /**
4124  * ext4_free_data - free a list of data blocks
4125  * @handle:     handle for this transaction
4126  * @inode:      inode we are dealing with
4127  * @this_bh:    indirect buffer_head which contains *@first and *@last
4128  * @first:      array of block numbers
4129  * @last:       points immediately past the end of array
4130  *
4131  * We are freeing all blocks refered from that array (numbers are stored as
4132  * little-endian 32-bit) and updating @inode->i_blocks appropriately.
4133  *
4134  * We accumulate contiguous runs of blocks to free.  Conveniently, if these
4135  * blocks are contiguous then releasing them at one time will only affect one
4136  * or two bitmap blocks (+ group descriptor(s) and superblock) and we won't
4137  * actually use a lot of journal space.
4138  *
4139  * @this_bh will be %NULL if @first and @last point into the inode's direct
4140  * block pointers.
4141  */
4142 static void ext4_free_data(handle_t *handle, struct inode *inode,
4143                            struct buffer_head *this_bh,
4144                            __le32 *first, __le32 *last)
4145 {
4146         ext4_fsblk_t block_to_free = 0;    /* Starting block # of a run */
4147         unsigned long count = 0;            /* Number of blocks in the run */
4148         __le32 *block_to_free_p = NULL;     /* Pointer into inode/ind
4149                                                corresponding to
4150                                                block_to_free */
4151         ext4_fsblk_t nr;                    /* Current block # */
4152         __le32 *p;                          /* Pointer into inode/ind
4153                                                for current block */
4154         int err;
4155
4156         if (this_bh) {                          /* For indirect block */
4157                 BUFFER_TRACE(this_bh, "get_write_access");
4158                 err = ext4_journal_get_write_access(handle, this_bh);
4159                 /* Important: if we can't update the indirect pointers
4160                  * to the blocks, we can't free them. */
4161                 if (err)
4162                         return;
4163         }
4164
4165         for (p = first; p < last; p++) {
4166                 nr = le32_to_cpu(*p);
4167                 if (nr) {
4168                         /* accumulate blocks to free if they're contiguous */
4169                         if (count == 0) {
4170                                 block_to_free = nr;
4171                                 block_to_free_p = p;
4172                                 count = 1;
4173                         } else if (nr == block_to_free + count) {
4174                                 count++;
4175                         } else {
4176                                 ext4_clear_blocks(handle, inode, this_bh,
4177                                                   block_to_free,
4178                                                   count, block_to_free_p, p);
4179                                 block_to_free = nr;
4180                                 block_to_free_p = p;
4181                                 count = 1;
4182                         }
4183                 }
4184         }
4185
4186         if (count > 0)
4187                 ext4_clear_blocks(handle, inode, this_bh, block_to_free,
4188                                   count, block_to_free_p, p);
4189
4190         if (this_bh) {
4191                 BUFFER_TRACE(this_bh, "call ext4_handle_dirty_metadata");
4192
4193                 /*
4194                  * The buffer head should have an attached journal head at this
4195                  * point. However, if the data is corrupted and an indirect
4196                  * block pointed to itself, it would have been detached when
4197                  * the block was cleared. Check for this instead of OOPSing.
4198                  */
4199                 if ((EXT4_JOURNAL(inode) == NULL) || bh2jh(this_bh))
4200                         ext4_handle_dirty_metadata(handle, inode, this_bh);
4201                 else
4202                         ext4_error(inode->i_sb, __func__,
4203                                    "circular indirect block detected, "
4204                                    "inode=%lu, block=%llu",
4205                                    inode->i_ino,
4206                                    (unsigned long long) this_bh->b_blocknr);
4207         }
4208 }
4209
4210 /**
4211  *      ext4_free_branches - free an array of branches
4212  *      @handle: JBD handle for this transaction
4213  *      @inode: inode we are dealing with
4214  *      @parent_bh: the buffer_head which contains *@first and *@last
4215  *      @first: array of block numbers
4216  *      @last:  pointer immediately past the end of array
4217  *      @depth: depth of the branches to free
4218  *
4219  *      We are freeing all blocks refered from these branches (numbers are
4220  *      stored as little-endian 32-bit) and updating @inode->i_blocks
4221  *      appropriately.
4222  */
4223 static void ext4_free_branches(handle_t *handle, struct inode *inode,
4224                                struct buffer_head *parent_bh,
4225                                __le32 *first, __le32 *last, int depth)
4226 {
4227         ext4_fsblk_t nr;
4228         __le32 *p;
4229
4230         if (ext4_handle_is_aborted(handle))
4231                 return;
4232
4233         if (depth--) {
4234                 struct buffer_head *bh;
4235                 int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4236                 p = last;
4237                 while (--p >= first) {
4238                         nr = le32_to_cpu(*p);
4239                         if (!nr)
4240                                 continue;               /* A hole */
4241
4242                         /* Go read the buffer for the next level down */
4243                         bh = sb_bread(inode->i_sb, nr);
4244
4245                         /*
4246                          * A read failure? Report error and clear slot
4247                          * (should be rare).
4248                          */
4249                         if (!bh) {
4250                                 ext4_error(inode->i_sb, "ext4_free_branches",
4251                                            "Read failure, inode=%lu, block=%llu",
4252                                            inode->i_ino, nr);
4253                                 continue;
4254                         }
4255
4256                         /* This zaps the entire block.  Bottom up. */
4257                         BUFFER_TRACE(bh, "free child branches");
4258                         ext4_free_branches(handle, inode, bh,
4259                                         (__le32 *) bh->b_data,
4260                                         (__le32 *) bh->b_data + addr_per_block,
4261                                         depth);
4262
4263                         /*
4264                          * We've probably journalled the indirect block several
4265                          * times during the truncate.  But it's no longer
4266                          * needed and we now drop it from the transaction via
4267                          * jbd2_journal_revoke().
4268                          *
4269                          * That's easy if it's exclusively part of this
4270                          * transaction.  But if it's part of the committing
4271                          * transaction then jbd2_journal_forget() will simply
4272                          * brelse() it.  That means that if the underlying
4273                          * block is reallocated in ext4_get_block(),
4274                          * unmap_underlying_metadata() will find this block
4275                          * and will try to get rid of it.  damn, damn.
4276                          *
4277                          * If this block has already been committed to the
4278                          * journal, a revoke record will be written.  And
4279                          * revoke records must be emitted *before* clearing
4280                          * this block's bit in the bitmaps.
4281                          */
4282                         ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
4283
4284                         /*
4285                          * Everything below this this pointer has been
4286                          * released.  Now let this top-of-subtree go.
4287                          *
4288                          * We want the freeing of this indirect block to be
4289                          * atomic in the journal with the updating of the
4290                          * bitmap block which owns it.  So make some room in
4291                          * the journal.
4292                          *
4293                          * We zero the parent pointer *after* freeing its
4294                          * pointee in the bitmaps, so if extend_transaction()
4295                          * for some reason fails to put the bitmap changes and
4296                          * the release into the same transaction, recovery
4297                          * will merely complain about releasing a free block,
4298                          * rather than leaking blocks.
4299                          */
4300                         if (ext4_handle_is_aborted(handle))
4301                                 return;
4302                         if (try_to_extend_transaction(handle, inode)) {
4303                                 ext4_mark_inode_dirty(handle, inode);
4304                                 ext4_truncate_restart_trans(handle, inode,
4305                                             blocks_for_truncate(inode));
4306                         }
4307
4308                         ext4_free_blocks(handle, inode, 0, nr, 1,
4309                                          EXT4_FREE_BLOCKS_METADATA);
4310
4311                         if (parent_bh) {
4312                                 /*
4313                                  * The block which we have just freed is
4314                                  * pointed to by an indirect block: journal it
4315                                  */
4316                                 BUFFER_TRACE(parent_bh, "get_write_access");
4317                                 if (!ext4_journal_get_write_access(handle,
4318                                                                    parent_bh)){
4319                                         *p = 0;
4320                                         BUFFER_TRACE(parent_bh,
4321                                         "call ext4_handle_dirty_metadata");
4322                                         ext4_handle_dirty_metadata(handle,
4323                                                                    inode,
4324                                                                    parent_bh);
4325                                 }
4326                         }
4327                 }
4328         } else {
4329                 /* We have reached the bottom of the tree. */
4330                 BUFFER_TRACE(parent_bh, "free data blocks");
4331                 ext4_free_data(handle, inode, parent_bh, first, last);
4332         }
4333 }
4334
4335 int ext4_can_truncate(struct inode *inode)
4336 {
4337         if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
4338                 return 0;
4339         if (S_ISREG(inode->i_mode))
4340                 return 1;
4341         if (S_ISDIR(inode->i_mode))
4342                 return 1;
4343         if (S_ISLNK(inode->i_mode))
4344                 return !ext4_inode_is_fast_symlink(inode);
4345         return 0;
4346 }
4347
4348 /*
4349  * ext4_truncate()
4350  *
4351  * We block out ext4_get_block() block instantiations across the entire
4352  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
4353  * simultaneously on behalf of the same inode.
4354  *
4355  * As we work through the truncate and commmit bits of it to the journal there
4356  * is one core, guiding principle: the file's tree must always be consistent on
4357  * disk.  We must be able to restart the truncate after a crash.
4358  *
4359  * The file's tree may be transiently inconsistent in memory (although it
4360  * probably isn't), but whenever we close off and commit a journal transaction,
4361  * the contents of (the filesystem + the journal) must be consistent and
4362  * restartable.  It's pretty simple, really: bottom up, right to left (although
4363  * left-to-right works OK too).
4364  *
4365  * Note that at recovery time, journal replay occurs *before* the restart of
4366  * truncate against the orphan inode list.
4367  *
4368  * The committed inode has the new, desired i_size (which is the same as
4369  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
4370  * that this inode's truncate did not complete and it will again call
4371  * ext4_truncate() to have another go.  So there will be instantiated blocks
4372  * to the right of the truncation point in a crashed ext4 filesystem.  But
4373  * that's fine - as long as they are linked from the inode, the post-crash
4374  * ext4_truncate() run will find them and release them.
4375  */
4376 void ext4_truncate(struct inode *inode)
4377 {
4378         handle_t *handle;
4379         struct ext4_inode_info *ei = EXT4_I(inode);
4380         __le32 *i_data = ei->i_data;
4381         int addr_per_block = EXT4_ADDR_PER_BLOCK(inode->i_sb);
4382         struct address_space *mapping = inode->i_mapping;
4383         ext4_lblk_t offsets[4];
4384         Indirect chain[4];
4385         Indirect *partial;
4386         __le32 nr = 0;
4387         int n;
4388         ext4_lblk_t last_block;
4389         unsigned blocksize = inode->i_sb->s_blocksize;
4390
4391         if (!ext4_can_truncate(inode))
4392                 return;
4393
4394         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
4395                 ei->i_state |= EXT4_STATE_DA_ALLOC_CLOSE;
4396
4397         if (EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL) {
4398                 ext4_ext_truncate(inode);
4399                 return;
4400         }
4401
4402         handle = start_transaction(inode);
4403         if (IS_ERR(handle))
4404                 return;         /* AKPM: return what? */
4405
4406         last_block = (inode->i_size + blocksize-1)
4407                                         >> EXT4_BLOCK_SIZE_BITS(inode->i_sb);
4408
4409         if (inode->i_size & (blocksize - 1))
4410                 if (ext4_block_truncate_page(handle, mapping, inode->i_size))
4411                         goto out_stop;
4412
4413         n = ext4_block_to_path(inode, last_block, offsets, NULL);
4414         if (n == 0)
4415                 goto out_stop;  /* error */
4416
4417         /*
4418          * OK.  This truncate is going to happen.  We add the inode to the
4419          * orphan list, so that if this truncate spans multiple transactions,
4420          * and we crash, we will resume the truncate when the filesystem
4421          * recovers.  It also marks the inode dirty, to catch the new size.
4422          *
4423          * Implication: the file must always be in a sane, consistent
4424          * truncatable state while each transaction commits.
4425          */
4426         if (ext4_orphan_add(handle, inode))
4427                 goto out_stop;
4428
4429         /*
4430          * From here we block out all ext4_get_block() callers who want to
4431          * modify the block allocation tree.
4432          */
4433         down_write(&ei->i_data_sem);
4434
4435         ext4_discard_preallocations(inode);
4436
4437         /*
4438          * The orphan list entry will now protect us from any crash which
4439          * occurs before the truncate completes, so it is now safe to propagate
4440          * the new, shorter inode size (held for now in i_size) into the
4441          * on-disk inode. We do this via i_disksize, which is the value which
4442          * ext4 *really* writes onto the disk inode.
4443          */
4444         ei->i_disksize = inode->i_size;
4445
4446         if (n == 1) {           /* direct blocks */
4447                 ext4_free_data(handle, inode, NULL, i_data+offsets[0],
4448                                i_data + EXT4_NDIR_BLOCKS);
4449                 goto do_indirects;
4450         }
4451
4452         partial = ext4_find_shared(inode, n, offsets, chain, &nr);
4453         /* Kill the top of shared branch (not detached) */
4454         if (nr) {
4455                 if (partial == chain) {
4456                         /* Shared branch grows from the inode */
4457                         ext4_free_branches(handle, inode, NULL,
4458                                            &nr, &nr+1, (chain+n-1) - partial);
4459                         *partial->p = 0;
4460                         /*
4461                          * We mark the inode dirty prior to restart,
4462                          * and prior to stop.  No need for it here.
4463                          */
4464                 } else {
4465                         /* Shared branch grows from an indirect block */
4466                         BUFFER_TRACE(partial->bh, "get_write_access");
4467                         ext4_free_branches(handle, inode, partial->bh,
4468                                         partial->p,
4469                                         partial->p+1, (chain+n-1) - partial);
4470                 }
4471         }
4472         /* Clear the ends of indirect blocks on the shared branch */
4473         while (partial > chain) {
4474                 ext4_free_branches(handle, inode, partial->bh, partial->p + 1,
4475                                    (__le32*)partial->bh->b_data+addr_per_block,
4476                                    (chain+n-1) - partial);
4477                 BUFFER_TRACE(partial->bh, "call brelse");
4478                 brelse(partial->bh);
4479                 partial--;
4480         }
4481 do_indirects:
4482         /* Kill the remaining (whole) subtrees */
4483         switch (offsets[0]) {
4484         default:
4485                 nr = i_data[EXT4_IND_BLOCK];
4486                 if (nr) {
4487                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 1);
4488                         i_data[EXT4_IND_BLOCK] = 0;
4489                 }
4490         case EXT4_IND_BLOCK:
4491                 nr = i_data[EXT4_DIND_BLOCK];
4492                 if (nr) {
4493                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 2);
4494                         i_data[EXT4_DIND_BLOCK] = 0;
4495                 }
4496         case EXT4_DIND_BLOCK:
4497                 nr = i_data[EXT4_TIND_BLOCK];
4498                 if (nr) {
4499                         ext4_free_branches(handle, inode, NULL, &nr, &nr+1, 3);
4500                         i_data[EXT4_TIND_BLOCK] = 0;
4501                 }
4502         case EXT4_TIND_BLOCK:
4503                 ;
4504         }
4505
4506         up_write(&ei->i_data_sem);
4507         inode->i_mtime = inode->i_ctime = ext4_current_time(inode);
4508         ext4_mark_inode_dirty(handle, inode);
4509
4510         /*
4511          * In a multi-transaction truncate, we only make the final transaction
4512          * synchronous
4513          */
4514         if (IS_SYNC(inode))
4515                 ext4_handle_sync(handle);
4516 out_stop:
4517         /*
4518          * If this was a simple ftruncate(), and the file will remain alive
4519          * then we need to clear up the orphan record which we created above.
4520          * However, if this was a real unlink then we were called by
4521          * ext4_delete_inode(), and we allow that function to clean up the
4522          * orphan info for us.
4523          */
4524         if (inode->i_nlink)
4525                 ext4_orphan_del(handle, inode);
4526
4527         ext4_journal_stop(handle);
4528 }
4529
4530 /*
4531  * ext4_get_inode_loc returns with an extra refcount against the inode's
4532  * underlying buffer_head on success. If 'in_mem' is true, we have all
4533  * data in memory that is needed to recreate the on-disk version of this
4534  * inode.
4535  */
4536 static int __ext4_get_inode_loc(struct inode *inode,
4537                                 struct ext4_iloc *iloc, int in_mem)
4538 {
4539         struct ext4_group_desc  *gdp;
4540         struct buffer_head      *bh;
4541         struct super_block      *sb = inode->i_sb;
4542         ext4_fsblk_t            block;
4543         int                     inodes_per_block, inode_offset;
4544
4545         iloc->bh = NULL;
4546         if (!ext4_valid_inum(sb, inode->i_ino))
4547                 return -EIO;
4548
4549         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
4550         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
4551         if (!gdp)
4552                 return -EIO;
4553
4554         /*
4555          * Figure out the offset within the block group inode table
4556          */
4557         inodes_per_block = (EXT4_BLOCK_SIZE(sb) / EXT4_INODE_SIZE(sb));
4558         inode_offset = ((inode->i_ino - 1) %
4559                         EXT4_INODES_PER_GROUP(sb));
4560         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
4561         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
4562
4563         bh = sb_getblk(sb, block);
4564         if (!bh) {
4565                 ext4_error(sb, "ext4_get_inode_loc", "unable to read "
4566                            "inode block - inode=%lu, block=%llu",
4567                            inode->i_ino, block);
4568                 return -EIO;
4569         }
4570         if (!buffer_uptodate(bh)) {
4571                 lock_buffer(bh);
4572
4573                 /*
4574                  * If the buffer has the write error flag, we have failed
4575                  * to write out another inode in the same block.  In this
4576                  * case, we don't have to read the block because we may
4577                  * read the old inode data successfully.
4578                  */
4579                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
4580                         set_buffer_uptodate(bh);
4581
4582                 if (buffer_uptodate(bh)) {
4583                         /* someone brought it uptodate while we waited */
4584                         unlock_buffer(bh);
4585                         goto has_buffer;
4586                 }
4587
4588                 /*
4589                  * If we have all information of the inode in memory and this
4590                  * is the only valid inode in the block, we need not read the
4591                  * block.
4592                  */
4593                 if (in_mem) {
4594                         struct buffer_head *bitmap_bh;
4595                         int i, start;
4596
4597                         start = inode_offset & ~(inodes_per_block - 1);
4598
4599                         /* Is the inode bitmap in cache? */
4600                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
4601                         if (!bitmap_bh)
4602                                 goto make_io;
4603
4604                         /*
4605                          * If the inode bitmap isn't in cache then the
4606                          * optimisation may end up performing two reads instead
4607                          * of one, so skip it.
4608                          */
4609                         if (!buffer_uptodate(bitmap_bh)) {
4610                                 brelse(bitmap_bh);
4611                                 goto make_io;
4612                         }
4613                         for (i = start; i < start + inodes_per_block; i++) {
4614                                 if (i == inode_offset)
4615                                         continue;
4616                                 if (ext4_test_bit(i, bitmap_bh->b_data))
4617                                         break;
4618                         }
4619                         brelse(bitmap_bh);
4620                         if (i == start + inodes_per_block) {
4621                                 /* all other inodes are free, so skip I/O */
4622                                 memset(bh->b_data, 0, bh->b_size);
4623                                 set_buffer_uptodate(bh);
4624                                 unlock_buffer(bh);
4625                                 goto has_buffer;
4626                         }
4627                 }
4628
4629 make_io:
4630                 /*
4631                  * If we need to do any I/O, try to pre-readahead extra
4632                  * blocks from the inode table.
4633                  */
4634                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
4635                         ext4_fsblk_t b, end, table;
4636                         unsigned num;
4637
4638                         table = ext4_inode_table(sb, gdp);
4639                         /* s_inode_readahead_blks is always a power of 2 */
4640                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
4641                         if (table > b)
4642                                 b = table;
4643                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
4644                         num = EXT4_INODES_PER_GROUP(sb);
4645                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4646                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
4647                                 num -= ext4_itable_unused_count(sb, gdp);
4648                         table += num / inodes_per_block;
4649                         if (end > table)
4650                                 end = table;
4651                         while (b <= end)
4652                                 sb_breadahead(sb, b++);
4653                 }
4654
4655                 /*
4656                  * There are other valid inodes in the buffer, this inode
4657                  * has in-inode xattrs, or we don't have this inode in memory.
4658                  * Read the block from disk.
4659                  */
4660                 get_bh(bh);
4661                 bh->b_end_io = end_buffer_read_sync;
4662                 submit_bh(READ_META, bh);
4663                 wait_on_buffer(bh);
4664                 if (!buffer_uptodate(bh)) {
4665                         ext4_error(sb, __func__,
4666                                    "unable to read inode block - inode=%lu, "
4667                                    "block=%llu", inode->i_ino, block);
4668                         brelse(bh);
4669                         return -EIO;
4670                 }
4671         }
4672 has_buffer:
4673         iloc->bh = bh;
4674         return 0;
4675 }
4676
4677 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
4678 {
4679         /* We have all inode data except xattrs in memory here. */
4680         return __ext4_get_inode_loc(inode, iloc,
4681                 !(EXT4_I(inode)->i_state & EXT4_STATE_XATTR));
4682 }
4683
4684 void ext4_set_inode_flags(struct inode *inode)
4685 {
4686         unsigned int flags = EXT4_I(inode)->i_flags;
4687
4688         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
4689         if (flags & EXT4_SYNC_FL)
4690                 inode->i_flags |= S_SYNC;
4691         if (flags & EXT4_APPEND_FL)
4692                 inode->i_flags |= S_APPEND;
4693         if (flags & EXT4_IMMUTABLE_FL)
4694                 inode->i_flags |= S_IMMUTABLE;
4695         if (flags & EXT4_NOATIME_FL)
4696                 inode->i_flags |= S_NOATIME;
4697         if (flags & EXT4_DIRSYNC_FL)
4698                 inode->i_flags |= S_DIRSYNC;
4699 }
4700
4701 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
4702 void ext4_get_inode_flags(struct ext4_inode_info *ei)
4703 {
4704         unsigned int flags = ei->vfs_inode.i_flags;
4705
4706         ei->i_flags &= ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
4707                         EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|EXT4_DIRSYNC_FL);
4708         if (flags & S_SYNC)
4709                 ei->i_flags |= EXT4_SYNC_FL;
4710         if (flags & S_APPEND)
4711                 ei->i_flags |= EXT4_APPEND_FL;
4712         if (flags & S_IMMUTABLE)
4713                 ei->i_flags |= EXT4_IMMUTABLE_FL;
4714         if (flags & S_NOATIME)
4715                 ei->i_flags |= EXT4_NOATIME_FL;
4716         if (flags & S_DIRSYNC)
4717                 ei->i_flags |= EXT4_DIRSYNC_FL;
4718 }
4719
4720 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
4721                                   struct ext4_inode_info *ei)
4722 {
4723         blkcnt_t i_blocks ;
4724         struct inode *inode = &(ei->vfs_inode);
4725         struct super_block *sb = inode->i_sb;
4726
4727         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
4728                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
4729                 /* we are using combined 48 bit field */
4730                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
4731                                         le32_to_cpu(raw_inode->i_blocks_lo);
4732                 if (ei->i_flags & EXT4_HUGE_FILE_FL) {
4733                         /* i_blocks represent file system block size */
4734                         return i_blocks  << (inode->i_blkbits - 9);
4735                 } else {
4736                         return i_blocks;
4737                 }
4738         } else {
4739                 return le32_to_cpu(raw_inode->i_blocks_lo);
4740         }
4741 }
4742
4743 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
4744 {
4745         struct ext4_iloc iloc;
4746         struct ext4_inode *raw_inode;
4747         struct ext4_inode_info *ei;
4748         struct inode *inode;
4749         journal_t *journal = EXT4_SB(sb)->s_journal;
4750         long ret;
4751         int block;
4752
4753         inode = iget_locked(sb, ino);
4754         if (!inode)
4755                 return ERR_PTR(-ENOMEM);
4756         if (!(inode->i_state & I_NEW))
4757                 return inode;
4758
4759         ei = EXT4_I(inode);
4760         iloc.bh = 0;
4761
4762         ret = __ext4_get_inode_loc(inode, &iloc, 0);
4763         if (ret < 0)
4764                 goto bad_inode;
4765         raw_inode = ext4_raw_inode(&iloc);
4766         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
4767         inode->i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
4768         inode->i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
4769         if (!(test_opt(inode->i_sb, NO_UID32))) {
4770                 inode->i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
4771                 inode->i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
4772         }
4773         inode->i_nlink = le16_to_cpu(raw_inode->i_links_count);
4774
4775         ei->i_state = 0;
4776         ei->i_dir_start_lookup = 0;
4777         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
4778         /* We now have enough fields to check if the inode was active or not.
4779          * This is needed because nfsd might try to access dead inodes
4780          * the test is that same one that e2fsck uses
4781          * NeilBrown 1999oct15
4782          */
4783         if (inode->i_nlink == 0) {
4784                 if (inode->i_mode == 0 ||
4785                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
4786                         /* this inode is deleted */
4787                         ret = -ESTALE;
4788                         goto bad_inode;
4789                 }
4790                 /* The only unlinked inodes we let through here have
4791                  * valid i_mode and are being read by the orphan
4792                  * recovery code: that's fine, we're about to complete
4793                  * the process of deleting those. */
4794         }
4795         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
4796         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
4797         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
4798         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
4799                 ei->i_file_acl |=
4800                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
4801         inode->i_size = ext4_isize(raw_inode);
4802         ei->i_disksize = inode->i_size;
4803 #ifdef CONFIG_QUOTA
4804         ei->i_reserved_quota = 0;
4805 #endif
4806         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
4807         ei->i_block_group = iloc.block_group;
4808         ei->i_last_alloc_group = ~0;
4809         /*
4810          * NOTE! The in-memory inode i_data array is in little-endian order
4811          * even on big-endian machines: we do NOT byteswap the block numbers!
4812          */
4813         for (block = 0; block < EXT4_N_BLOCKS; block++)
4814                 ei->i_data[block] = raw_inode->i_block[block];
4815         INIT_LIST_HEAD(&ei->i_orphan);
4816
4817         /*
4818          * Set transaction id's of transactions that have to be committed
4819          * to finish f[data]sync. We set them to currently running transaction
4820          * as we cannot be sure that the inode or some of its metadata isn't
4821          * part of the transaction - the inode could have been reclaimed and
4822          * now it is reread from disk.
4823          */
4824         if (journal) {
4825                 transaction_t *transaction;
4826                 tid_t tid;
4827
4828                 spin_lock(&journal->j_state_lock);
4829                 if (journal->j_running_transaction)
4830                         transaction = journal->j_running_transaction;
4831                 else
4832                         transaction = journal->j_committing_transaction;
4833                 if (transaction)
4834                         tid = transaction->t_tid;
4835                 else
4836                         tid = journal->j_commit_sequence;
4837                 spin_unlock(&journal->j_state_lock);
4838                 ei->i_sync_tid = tid;
4839                 ei->i_datasync_tid = tid;
4840         }
4841
4842         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4843                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
4844                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
4845                     EXT4_INODE_SIZE(inode->i_sb)) {
4846                         ret = -EIO;
4847                         goto bad_inode;
4848                 }
4849                 if (ei->i_extra_isize == 0) {
4850                         /* The extra space is currently unused. Use it. */
4851                         ei->i_extra_isize = sizeof(struct ext4_inode) -
4852                                             EXT4_GOOD_OLD_INODE_SIZE;
4853                 } else {
4854                         __le32 *magic = (void *)raw_inode +
4855                                         EXT4_GOOD_OLD_INODE_SIZE +
4856                                         ei->i_extra_isize;
4857                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
4858                                 ei->i_state |= EXT4_STATE_XATTR;
4859                 }
4860         } else
4861                 ei->i_extra_isize = 0;
4862
4863         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
4864         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
4865         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
4866         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
4867
4868         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
4869         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
4870                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
4871                         inode->i_version |=
4872                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
4873         }
4874
4875         ret = 0;
4876         if (ei->i_file_acl &&
4877             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
4878                 ext4_error(sb, __func__,
4879                            "bad extended attribute block %llu in inode #%lu",
4880                            ei->i_file_acl, inode->i_ino);
4881                 ret = -EIO;
4882                 goto bad_inode;
4883         } else if (ei->i_flags & EXT4_EXTENTS_FL) {
4884                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4885                     (S_ISLNK(inode->i_mode) &&
4886                      !ext4_inode_is_fast_symlink(inode)))
4887                         /* Validate extent which is part of inode */
4888                         ret = ext4_ext_check_inode(inode);
4889         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
4890                    (S_ISLNK(inode->i_mode) &&
4891                     !ext4_inode_is_fast_symlink(inode))) {
4892                 /* Validate block references which are part of inode */
4893                 ret = ext4_check_inode_blockref(inode);
4894         }
4895         if (ret)
4896                 goto bad_inode;
4897
4898         if (S_ISREG(inode->i_mode)) {
4899                 inode->i_op = &ext4_file_inode_operations;
4900                 inode->i_fop = &ext4_file_operations;
4901                 ext4_set_aops(inode);
4902         } else if (S_ISDIR(inode->i_mode)) {
4903                 inode->i_op = &ext4_dir_inode_operations;
4904                 inode->i_fop = &ext4_dir_operations;
4905         } else if (S_ISLNK(inode->i_mode)) {
4906                 if (ext4_inode_is_fast_symlink(inode)) {
4907                         inode->i_op = &ext4_fast_symlink_inode_operations;
4908                         nd_terminate_link(ei->i_data, inode->i_size,
4909                                 sizeof(ei->i_data) - 1);
4910                 } else {
4911                         inode->i_op = &ext4_symlink_inode_operations;
4912                         ext4_set_aops(inode);
4913                 }
4914         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
4915               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
4916                 inode->i_op = &ext4_special_inode_operations;
4917                 if (raw_inode->i_block[0])
4918                         init_special_inode(inode, inode->i_mode,
4919                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
4920                 else
4921                         init_special_inode(inode, inode->i_mode,
4922                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
4923         } else {
4924                 ret = -EIO;
4925                 ext4_error(inode->i_sb, __func__,
4926                            "bogus i_mode (%o) for inode=%lu",
4927                            inode->i_mode, inode->i_ino);
4928                 goto bad_inode;
4929         }
4930         brelse(iloc.bh);
4931         ext4_set_inode_flags(inode);
4932         unlock_new_inode(inode);
4933         return inode;
4934
4935 bad_inode:
4936         brelse(iloc.bh);
4937         iget_failed(inode);
4938         return ERR_PTR(ret);
4939 }
4940
4941 static int ext4_inode_blocks_set(handle_t *handle,
4942                                 struct ext4_inode *raw_inode,
4943                                 struct ext4_inode_info *ei)
4944 {
4945         struct inode *inode = &(ei->vfs_inode);
4946         u64 i_blocks = inode->i_blocks;
4947         struct super_block *sb = inode->i_sb;
4948
4949         if (i_blocks <= ~0U) {
4950                 /*
4951                  * i_blocks can be represnted in a 32 bit variable
4952                  * as multiple of 512 bytes
4953                  */
4954                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4955                 raw_inode->i_blocks_high = 0;
4956                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4957                 return 0;
4958         }
4959         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
4960                 return -EFBIG;
4961
4962         if (i_blocks <= 0xffffffffffffULL) {
4963                 /*
4964                  * i_blocks can be represented in a 48 bit variable
4965                  * as multiple of 512 bytes
4966                  */
4967                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4968                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4969                 ei->i_flags &= ~EXT4_HUGE_FILE_FL;
4970         } else {
4971                 ei->i_flags |= EXT4_HUGE_FILE_FL;
4972                 /* i_block is stored in file system block size */
4973                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
4974                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
4975                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
4976         }
4977         return 0;
4978 }
4979
4980 /*
4981  * Post the struct inode info into an on-disk inode location in the
4982  * buffer-cache.  This gobbles the caller's reference to the
4983  * buffer_head in the inode location struct.
4984  *
4985  * The caller must have write access to iloc->bh.
4986  */
4987 static int ext4_do_update_inode(handle_t *handle,
4988                                 struct inode *inode,
4989                                 struct ext4_iloc *iloc)
4990 {
4991         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
4992         struct ext4_inode_info *ei = EXT4_I(inode);
4993         struct buffer_head *bh = iloc->bh;
4994         int err = 0, rc, block;
4995
4996         /* For fields not not tracking in the in-memory inode,
4997          * initialise them to zero for new inodes. */
4998         if (ei->i_state & EXT4_STATE_NEW)
4999                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
5000
5001         ext4_get_inode_flags(ei);
5002         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
5003         if (!(test_opt(inode->i_sb, NO_UID32))) {
5004                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(inode->i_uid));
5005                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(inode->i_gid));
5006 /*
5007  * Fix up interoperability with old kernels. Otherwise, old inodes get
5008  * re-used with the upper 16 bits of the uid/gid intact
5009  */
5010                 if (!ei->i_dtime) {
5011                         raw_inode->i_uid_high =
5012                                 cpu_to_le16(high_16_bits(inode->i_uid));
5013                         raw_inode->i_gid_high =
5014                                 cpu_to_le16(high_16_bits(inode->i_gid));
5015                 } else {
5016                         raw_inode->i_uid_high = 0;
5017                         raw_inode->i_gid_high = 0;
5018                 }
5019         } else {
5020                 raw_inode->i_uid_low =
5021                         cpu_to_le16(fs_high2lowuid(inode->i_uid));
5022                 raw_inode->i_gid_low =
5023                         cpu_to_le16(fs_high2lowgid(inode->i_gid));
5024                 raw_inode->i_uid_high = 0;
5025                 raw_inode->i_gid_high = 0;
5026         }
5027         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
5028
5029         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
5030         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
5031         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
5032         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
5033
5034         if (ext4_inode_blocks_set(handle, raw_inode, ei))
5035                 goto out_brelse;
5036         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
5037         raw_inode->i_flags = cpu_to_le32(ei->i_flags);
5038         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
5039             cpu_to_le32(EXT4_OS_HURD))
5040                 raw_inode->i_file_acl_high =
5041                         cpu_to_le16(ei->i_file_acl >> 32);
5042         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
5043         ext4_isize_set(raw_inode, ei->i_disksize);
5044         if (ei->i_disksize > 0x7fffffffULL) {
5045                 struct super_block *sb = inode->i_sb;
5046                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
5047                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
5048                                 EXT4_SB(sb)->s_es->s_rev_level ==
5049                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
5050                         /* If this is the first large file
5051                          * created, add a flag to the superblock.
5052                          */
5053                         err = ext4_journal_get_write_access(handle,
5054                                         EXT4_SB(sb)->s_sbh);
5055                         if (err)
5056                                 goto out_brelse;
5057                         ext4_update_dynamic_rev(sb);
5058                         EXT4_SET_RO_COMPAT_FEATURE(sb,
5059                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
5060                         sb->s_dirt = 1;
5061                         ext4_handle_sync(handle);
5062                         err = ext4_handle_dirty_metadata(handle, inode,
5063                                         EXT4_SB(sb)->s_sbh);
5064                 }
5065         }
5066         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
5067         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
5068                 if (old_valid_dev(inode->i_rdev)) {
5069                         raw_inode->i_block[0] =
5070                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
5071                         raw_inode->i_block[1] = 0;
5072                 } else {
5073                         raw_inode->i_block[0] = 0;
5074                         raw_inode->i_block[1] =
5075                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
5076                         raw_inode->i_block[2] = 0;
5077                 }
5078         } else
5079                 for (block = 0; block < EXT4_N_BLOCKS; block++)
5080                         raw_inode->i_block[block] = ei->i_data[block];
5081
5082         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
5083         if (ei->i_extra_isize) {
5084                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
5085                         raw_inode->i_version_hi =
5086                         cpu_to_le32(inode->i_version >> 32);
5087                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
5088         }
5089
5090         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
5091         rc = ext4_handle_dirty_metadata(handle, inode, bh);
5092         if (!err)
5093                 err = rc;
5094         ei->i_state &= ~EXT4_STATE_NEW;
5095
5096         ext4_update_inode_fsync_trans(handle, inode, 0);
5097 out_brelse:
5098         brelse(bh);
5099         ext4_std_error(inode->i_sb, err);
5100         return err;
5101 }
5102
5103 /*
5104  * ext4_write_inode()
5105  *
5106  * We are called from a few places:
5107  *
5108  * - Within generic_file_write() for O_SYNC files.
5109  *   Here, there will be no transaction running. We wait for any running
5110  *   trasnaction to commit.
5111  *
5112  * - Within sys_sync(), kupdate and such.
5113  *   We wait on commit, if tol to.
5114  *
5115  * - Within prune_icache() (PF_MEMALLOC == true)
5116  *   Here we simply return.  We can't afford to block kswapd on the
5117  *   journal commit.
5118  *
5119  * In all cases it is actually safe for us to return without doing anything,
5120  * because the inode has been copied into a raw inode buffer in
5121  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
5122  * knfsd.
5123  *
5124  * Note that we are absolutely dependent upon all inode dirtiers doing the
5125  * right thing: they *must* call mark_inode_dirty() after dirtying info in
5126  * which we are interested.
5127  *
5128  * It would be a bug for them to not do this.  The code:
5129  *
5130  *      mark_inode_dirty(inode)
5131  *      stuff();
5132  *      inode->i_size = expr;
5133  *
5134  * is in error because a kswapd-driven write_inode() could occur while
5135  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
5136  * will no longer be on the superblock's dirty inode list.
5137  */
5138 int ext4_write_inode(struct inode *inode, int wait)
5139 {
5140         int err;
5141
5142         if (current->flags & PF_MEMALLOC)
5143                 return 0;
5144
5145         if (EXT4_SB(inode->i_sb)->s_journal) {
5146                 if (ext4_journal_current_handle()) {
5147                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
5148                         dump_stack();
5149                         return -EIO;
5150                 }
5151
5152                 if (!wait)
5153                         return 0;
5154
5155                 err = ext4_force_commit(inode->i_sb);
5156         } else {
5157                 struct ext4_iloc iloc;
5158
5159                 err = ext4_get_inode_loc(inode, &iloc);
5160                 if (err)
5161                         return err;
5162                 if (wait)
5163                         sync_dirty_buffer(iloc.bh);
5164                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
5165                         ext4_error(inode->i_sb, __func__,
5166                                    "IO error syncing inode, "
5167                                    "inode=%lu, block=%llu",
5168                                    inode->i_ino,
5169                                    (unsigned long long)iloc.bh->b_blocknr);
5170                         err = -EIO;
5171                 }
5172         }
5173         return err;
5174 }
5175
5176 /*
5177  * ext4_setattr()
5178  *
5179  * Called from notify_change.
5180  *
5181  * We want to trap VFS attempts to truncate the file as soon as
5182  * possible.  In particular, we want to make sure that when the VFS
5183  * shrinks i_size, we put the inode on the orphan list and modify
5184  * i_disksize immediately, so that during the subsequent flushing of
5185  * dirty pages and freeing of disk blocks, we can guarantee that any
5186  * commit will leave the blocks being flushed in an unused state on
5187  * disk.  (On recovery, the inode will get truncated and the blocks will
5188  * be freed, so we have a strong guarantee that no future commit will
5189  * leave these blocks visible to the user.)
5190  *
5191  * Another thing we have to assure is that if we are in ordered mode
5192  * and inode is still attached to the committing transaction, we must
5193  * we start writeout of all the dirty pages which are being truncated.
5194  * This way we are sure that all the data written in the previous
5195  * transaction are already on disk (truncate waits for pages under
5196  * writeback).
5197  *
5198  * Called with inode->i_mutex down.
5199  */
5200 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
5201 {
5202         struct inode *inode = dentry->d_inode;
5203         int error, rc = 0;
5204         const unsigned int ia_valid = attr->ia_valid;
5205
5206         error = inode_change_ok(inode, attr);
5207         if (error)
5208                 return error;
5209
5210         if ((ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
5211                 (ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
5212                 handle_t *handle;
5213
5214                 /* (user+group)*(old+new) structure, inode write (sb,
5215                  * inode block, ? - but truncate inode update has it) */
5216                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
5217                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
5218                 if (IS_ERR(handle)) {
5219                         error = PTR_ERR(handle);
5220                         goto err_out;
5221                 }
5222                 error = vfs_dq_transfer(inode, attr) ? -EDQUOT : 0;
5223                 if (error) {
5224                         ext4_journal_stop(handle);
5225                         return error;
5226                 }
5227                 /* Update corresponding info in inode so that everything is in
5228                  * one transaction */
5229                 if (attr->ia_valid & ATTR_UID)
5230                         inode->i_uid = attr->ia_uid;
5231                 if (attr->ia_valid & ATTR_GID)
5232                         inode->i_gid = attr->ia_gid;
5233                 error = ext4_mark_inode_dirty(handle, inode);
5234                 ext4_journal_stop(handle);
5235         }
5236
5237         if (attr->ia_valid & ATTR_SIZE) {
5238                 if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL)) {
5239                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5240
5241                         if (attr->ia_size > sbi->s_bitmap_maxbytes) {
5242                                 error = -EFBIG;
5243                                 goto err_out;
5244                         }
5245                 }
5246         }
5247
5248         if (S_ISREG(inode->i_mode) &&
5249             attr->ia_valid & ATTR_SIZE && attr->ia_size < inode->i_size) {
5250                 handle_t *handle;
5251
5252                 handle = ext4_journal_start(inode, 3);
5253                 if (IS_ERR(handle)) {
5254                         error = PTR_ERR(handle);
5255                         goto err_out;
5256                 }
5257
5258                 error = ext4_orphan_add(handle, inode);
5259                 EXT4_I(inode)->i_disksize = attr->ia_size;
5260                 rc = ext4_mark_inode_dirty(handle, inode);
5261                 if (!error)
5262                         error = rc;
5263                 ext4_journal_stop(handle);
5264
5265                 if (ext4_should_order_data(inode)) {
5266                         error = ext4_begin_ordered_truncate(inode,
5267                                                             attr->ia_size);
5268                         if (error) {
5269                                 /* Do as much error cleanup as possible */
5270                                 handle = ext4_journal_start(inode, 3);
5271                                 if (IS_ERR(handle)) {
5272                                         ext4_orphan_del(NULL, inode);
5273                                         goto err_out;
5274                                 }
5275                                 ext4_orphan_del(handle, inode);
5276                                 ext4_journal_stop(handle);
5277                                 goto err_out;
5278                         }
5279                 }
5280         }
5281
5282         rc = inode_setattr(inode, attr);
5283
5284         /* If inode_setattr's call to ext4_truncate failed to get a
5285          * transaction handle at all, we need to clean up the in-core
5286          * orphan list manually. */
5287         if (inode->i_nlink)
5288                 ext4_orphan_del(NULL, inode);
5289
5290         if (!rc && (ia_valid & ATTR_MODE))
5291                 rc = ext4_acl_chmod(inode);
5292
5293 err_out:
5294         ext4_std_error(inode->i_sb, error);
5295         if (!error)
5296                 error = rc;
5297         return error;
5298 }
5299
5300 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
5301                  struct kstat *stat)
5302 {
5303         struct inode *inode;
5304         unsigned long delalloc_blocks;
5305
5306         inode = dentry->d_inode;
5307         generic_fillattr(inode, stat);
5308
5309         /*
5310          * We can't update i_blocks if the block allocation is delayed
5311          * otherwise in the case of system crash before the real block
5312          * allocation is done, we will have i_blocks inconsistent with
5313          * on-disk file blocks.
5314          * We always keep i_blocks updated together with real
5315          * allocation. But to not confuse with user, stat
5316          * will return the blocks that include the delayed allocation
5317          * blocks for this file.
5318          */
5319         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
5320         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
5321         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
5322
5323         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
5324         return 0;
5325 }
5326
5327 static int ext4_indirect_trans_blocks(struct inode *inode, int nrblocks,
5328                                       int chunk)
5329 {
5330         int indirects;
5331
5332         /* if nrblocks are contiguous */
5333         if (chunk) {
5334                 /*
5335                  * With N contiguous data blocks, it need at most
5336                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) indirect blocks
5337                  * 2 dindirect blocks
5338                  * 1 tindirect block
5339                  */
5340                 indirects = nrblocks / EXT4_ADDR_PER_BLOCK(inode->i_sb);
5341                 return indirects + 3;
5342         }
5343         /*
5344          * if nrblocks are not contiguous, worse case, each block touch
5345          * a indirect block, and each indirect block touch a double indirect
5346          * block, plus a triple indirect block
5347          */
5348         indirects = nrblocks * 2 + 1;
5349         return indirects;
5350 }
5351
5352 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5353 {
5354         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
5355                 return ext4_indirect_trans_blocks(inode, nrblocks, chunk);
5356         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
5357 }
5358
5359 /*
5360  * Account for index blocks, block groups bitmaps and block group
5361  * descriptor blocks if modify datablocks and index blocks
5362  * worse case, the indexs blocks spread over different block groups
5363  *
5364  * If datablocks are discontiguous, they are possible to spread over
5365  * different block groups too. If they are contiuguous, with flexbg,
5366  * they could still across block group boundary.
5367  *
5368  * Also account for superblock, inode, quota and xattr blocks
5369  */
5370 int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
5371 {
5372         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
5373         int gdpblocks;
5374         int idxblocks;
5375         int ret = 0;
5376
5377         /*
5378          * How many index blocks need to touch to modify nrblocks?
5379          * The "Chunk" flag indicating whether the nrblocks is
5380          * physically contiguous on disk
5381          *
5382          * For Direct IO and fallocate, they calls get_block to allocate
5383          * one single extent at a time, so they could set the "Chunk" flag
5384          */
5385         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
5386
5387         ret = idxblocks;
5388
5389         /*
5390          * Now let's see how many group bitmaps and group descriptors need
5391          * to account
5392          */
5393         groups = idxblocks;
5394         if (chunk)
5395                 groups += 1;
5396         else
5397                 groups += nrblocks;
5398
5399         gdpblocks = groups;
5400         if (groups > ngroups)
5401                 groups = ngroups;
5402         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
5403                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
5404
5405         /* bitmaps and block group descriptor blocks */
5406         ret += groups + gdpblocks;
5407
5408         /* Blocks for super block, inode, quota and xattr blocks */
5409         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
5410
5411         return ret;
5412 }
5413
5414 /*
5415  * Calulate the total number of credits to reserve to fit
5416  * the modification of a single pages into a single transaction,
5417  * which may include multiple chunks of block allocations.
5418  *
5419  * This could be called via ext4_write_begin()
5420  *
5421  * We need to consider the worse case, when
5422  * one new block per extent.
5423  */
5424 int ext4_writepage_trans_blocks(struct inode *inode)
5425 {
5426         int bpp = ext4_journal_blocks_per_page(inode);
5427         int ret;
5428
5429         ret = ext4_meta_trans_blocks(inode, bpp, 0);
5430
5431         /* Account for data blocks for journalled mode */
5432         if (ext4_should_journal_data(inode))
5433                 ret += bpp;
5434         return ret;
5435 }
5436
5437 /*
5438  * Calculate the journal credits for a chunk of data modification.
5439  *
5440  * This is called from DIO, fallocate or whoever calling
5441  * ext4_get_blocks() to map/allocate a chunk of contiguous disk blocks.
5442  *
5443  * journal buffers for data blocks are not included here, as DIO
5444  * and fallocate do no need to journal data buffers.
5445  */
5446 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
5447 {
5448         return ext4_meta_trans_blocks(inode, nrblocks, 1);
5449 }
5450
5451 /*
5452  * The caller must have previously called ext4_reserve_inode_write().
5453  * Give this, we know that the caller already has write access to iloc->bh.
5454  */
5455 int ext4_mark_iloc_dirty(handle_t *handle,
5456                          struct inode *inode, struct ext4_iloc *iloc)
5457 {
5458         int err = 0;
5459
5460         if (test_opt(inode->i_sb, I_VERSION))
5461                 inode_inc_iversion(inode);
5462
5463         /* the do_update_inode consumes one bh->b_count */
5464         get_bh(iloc->bh);
5465
5466         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
5467         err = ext4_do_update_inode(handle, inode, iloc);
5468         put_bh(iloc->bh);
5469         return err;
5470 }
5471
5472 /*
5473  * On success, We end up with an outstanding reference count against
5474  * iloc->bh.  This _must_ be cleaned up later.
5475  */
5476
5477 int
5478 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
5479                          struct ext4_iloc *iloc)
5480 {
5481         int err;
5482
5483         err = ext4_get_inode_loc(inode, iloc);
5484         if (!err) {
5485                 BUFFER_TRACE(iloc->bh, "get_write_access");
5486                 err = ext4_journal_get_write_access(handle, iloc->bh);
5487                 if (err) {
5488                         brelse(iloc->bh);
5489                         iloc->bh = NULL;
5490                 }
5491         }
5492         ext4_std_error(inode->i_sb, err);
5493         return err;
5494 }
5495
5496 /*
5497  * Expand an inode by new_extra_isize bytes.
5498  * Returns 0 on success or negative error number on failure.
5499  */
5500 static int ext4_expand_extra_isize(struct inode *inode,
5501                                    unsigned int new_extra_isize,
5502                                    struct ext4_iloc iloc,
5503                                    handle_t *handle)
5504 {
5505         struct ext4_inode *raw_inode;
5506         struct ext4_xattr_ibody_header *header;
5507         struct ext4_xattr_entry *entry;
5508
5509         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
5510                 return 0;
5511
5512         raw_inode = ext4_raw_inode(&iloc);
5513
5514         header = IHDR(inode, raw_inode);
5515         entry = IFIRST(header);
5516
5517         /* No extended attributes present */
5518         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR) ||
5519                 header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
5520                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
5521                         new_extra_isize);
5522                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
5523                 return 0;
5524         }
5525
5526         /* try to expand with EAs present */
5527         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
5528                                           raw_inode, handle);
5529 }
5530
5531 /*
5532  * What we do here is to mark the in-core inode as clean with respect to inode
5533  * dirtiness (it may still be data-dirty).
5534  * This means that the in-core inode may be reaped by prune_icache
5535  * without having to perform any I/O.  This is a very good thing,
5536  * because *any* task may call prune_icache - even ones which
5537  * have a transaction open against a different journal.
5538  *
5539  * Is this cheating?  Not really.  Sure, we haven't written the
5540  * inode out, but prune_icache isn't a user-visible syncing function.
5541  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
5542  * we start and wait on commits.
5543  *
5544  * Is this efficient/effective?  Well, we're being nice to the system
5545  * by cleaning up our inodes proactively so they can be reaped
5546  * without I/O.  But we are potentially leaving up to five seconds'
5547  * worth of inodes floating about which prune_icache wants us to
5548  * write out.  One way to fix that would be to get prune_icache()
5549  * to do a write_super() to free up some memory.  It has the desired
5550  * effect.
5551  */
5552 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
5553 {
5554         struct ext4_iloc iloc;
5555         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
5556         static unsigned int mnt_count;
5557         int err, ret;
5558
5559         might_sleep();
5560         err = ext4_reserve_inode_write(handle, inode, &iloc);
5561         if (ext4_handle_valid(handle) &&
5562             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
5563             !(EXT4_I(inode)->i_state & EXT4_STATE_NO_EXPAND)) {
5564                 /*
5565                  * We need extra buffer credits since we may write into EA block
5566                  * with this same handle. If journal_extend fails, then it will
5567                  * only result in a minor loss of functionality for that inode.
5568                  * If this is felt to be critical, then e2fsck should be run to
5569                  * force a large enough s_min_extra_isize.
5570                  */
5571                 if ((jbd2_journal_extend(handle,
5572                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
5573                         ret = ext4_expand_extra_isize(inode,
5574                                                       sbi->s_want_extra_isize,
5575                                                       iloc, handle);
5576                         if (ret) {
5577                                 EXT4_I(inode)->i_state |= EXT4_STATE_NO_EXPAND;
5578                                 if (mnt_count !=
5579                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
5580                                         ext4_warning(inode->i_sb, __func__,
5581                                         "Unable to expand inode %lu. Delete"
5582                                         " some EAs or run e2fsck.",
5583                                         inode->i_ino);
5584                                         mnt_count =
5585                                           le16_to_cpu(sbi->s_es->s_mnt_count);
5586                                 }
5587                         }
5588                 }
5589         }
5590         if (!err)
5591                 err